JohnnyCode.ai
Back to homePlain English

Model Guide

Plain-English notes on what these models are good and bad at. Benchmarks are secondary — if we do not have a verified number, we say so.

Each card is a short take: what the model is for, where it struggles, when to reach for it, peers, and local fit for NVIDIA 12GB VRAM and MacBook M4 Pro Max 36GB unified memory.

Verdict labels: fits · fits_quantized · tight · no · n_a. We never invent exact VRAM numbers.

Best Model Index

Which model is best overall?

One score per model from Artificial Analysis data: how smart it is, what a task costs, and how fast it answers. Cloud and local are ranked separately because they answer different questions. Move the weights to match what you care about.

Captured 2026-10-15. Coding cross-check from World of AI Bench (public top models, 2026-09-23).

Cloud: best overall

Every model you can call through an API, frontier labs and hosted open weights together. Each model is shown at its best-scoring reasoning setting.

Proprietary (frontier labs)Open weights, hosted
  1. K2 Horizon 375B A23B71.8
  2. Claude Haiku 5.5 max69.9
  3. Claude Sonnet 5.5 max69.1
  4. Claude Opus 5.5 max68.8
  5. GPT-6.1 Sol xhigh68.1
  6. Muse Wonder max66.9
  7. MiMo-V2.6-Pro66.5
  8. Muse Spark 1.3 xhigh66.2
  9. GPT-6 Luna max64.7
  10. DeepSeek V4.1 Flash max63.5
  11. GPT-6 Astra xhigh62.7
  12. Claude Fable 5.1 xhigh62.2
  13. Ling 3.1 Flash61.7
  14. MiMo-V2.6-Flash61.0
  15. GLM-5.3-Flash60.8
Show as table
#ModelCreatorWeightsScoreIntel. IndexCost/taskTokens/sWoAI coding
1K2 Horizon 375B A23B Institute of Foundation ModelsOpen71.830.5<$0.01891—
2Claude Haiku 5.5 maxAnthropicProprietary69.943.4$0.21238—
3Claude Sonnet 5.5 maxAnthropicProprietary69.156.0$5.46133—
4Claude Opus 5.5 maxAnthropicProprietary68.857.6$5.989688.0
5GPT-6.1 Sol xhighOpenAIProprietary68.151.0$0.3951—
6Muse Wonder maxMetaProprietary66.952.4$4.52161—
7MiMo-V2.6-Pro XiaomiOpen66.546.3$0.134783.3
8Muse Spark 1.3 xhighMetaProprietary66.245.1$1.37283—
9GPT-6 Luna maxOpenAIProprietary64.738.1$0.0712081.0
10DeepSeek V4.1 Flash maxDeepSeekOpen63.539.5$0.2717781.8
11GPT-6 Astra xhighOpenAIProprietary62.752.4$2.314487.7
12Claude Fable 5.1 xhighAnthropicProprietary62.253.2$5.986386.5
13Ling 3.1 Flash InclusionAIProprietary61.741.1$1.00213—
14MiMo-V2.6-Flash XiaomiOpen61.037.9$0.0654—
15GLM-5.3-Flash Z AIOpen60.841.8$0.2865—

Local: best you can run yourself

Open-weight models up to about 40B parameters. Cost and API speed don't apply at home, so the score trades intelligence against memory: smaller runs faster and leaves room for context.

  1. Qwen3.8 27B xhigh49.5 ~18 GB
  2. K2 Horizon 7B39.2 ~6 GB
  3. K2 Horizon MoVA 36B A4B36.0 ~23 GB
  4. K2 Horizon 3.7B35.3 ~4 GB
  5. MiniCPM5-2B33.2 ~3 GB
  6. Qwen3.5 4B reasoning31.5 ~4 GB
  7. G9v3-39A5B30.6 ~25 GB
  8. MiniCPM5-1B reasoning29.9 ~2 GB
  9. G9v3-3B29.5 ~3 GB
  10. Qwen3.5 9B non-reasoning27.7 ~7 GB
  11. Gemma 4 12B reasoning27.4 ~9 GB
  12. Granite 4.2 3B27.1 ~3 GB
  13. LFM2.5-2.6B26.7 ~3 GB
  14. Qwen3.5 0.8B reasoning26.6 ~2 GB
  15. Gemma 3 270M26.4 ~2 GB

Fit is estimated at 4-bit quantization (Q4_K_M or MLX 4-bit): about 0.6 GB per billion parameters plus 1.5 GB. "Offload" is a mixture-of-experts model that runs on 12GB by keeping inactive experts in system RAM; slower, but usable.

Show as table
#ModelCreatorScoreIntel. IndexParamsGB at 4-bit12GB GPU36GB Mac
1Qwen3.8 27B xhighAlibaba49.533.727B17.7nofits
2K2 Horizon 7B Institute of Foundation Models39.220.67B5.7fitsfits
3K2 Horizon MoVA 36B A4B Institute of Foundation Models36.025.336B / 4B23.1offloadfits
4K2 Horizon 3.7B Institute of Foundation Models35.315.63.7B3.7fitsfits
5MiniCPM5-2B OpenBMB33.212.52B2.7fitsfits
6Qwen3.5 4B reasoningAlibaba31.513.1 est.4B3.9fitsfits
7G9v3-39A5B AI9Stars30.621.8 est.39B / 5B24.9offloadfits
8MiniCPM5-1B reasoningOpenBMB29.98.8 est.1B2.1fitsfits
9G9v3-3B AI9Stars29.510.8 est.3B3.3fitsfits
10Qwen3.5 9B non-reasoningAlibaba27.713.3 est.9B6.9fitsfits
11Gemma 4 12B reasoningGoogle27.414.2 est.12B8.7fitsfits
12Granite 4.2 3B IBM27.19.13B3.3fitsfits
13LFM2.5-2.6B Liquid AI26.78.4 est.2.6B3.1fitsfits
14Qwen3.5 0.8B reasoningAlibaba26.66.1 est.0.8B2.0fitsfits
15Gemma 3 270M Google26.45.1 est.0.27B1.7fitsfits

How the score works

Intelligence
Artificial Analysis Intelligence Index divided by the top score. It already blends reasoning, knowledge, coding and agentic evals.
Cost
What a task actually cost to run, reasoning tokens included. Log scale from $0.005 (full marks) to $8 (zero), so halving cost always earns the same.
Speed
Median output tokens per second, log scale from 20 (zero) to 400 (full marks).
Composite
100 × the weighted average of the parts. Cloud models need measured cost and speed to be ranked; "est." marks estimated intelligence scores.

Specialty models

Best at one thing

Voice, music, images, video, 3D, world models, documents and search each have their own leaderboard, or none at all. Where one exists, the best frontier models you reach through an API sit next to the best open weights you can run yourself. Below each ranking are strong open models the leaderboard hasn't scored yet, found on Hugging Face and checked by hand.

Open weights doesn't mean small: check the size and the 12GB / 36GB fit before planning to run a model on your own machine. Captured 2026-11-07.

Text to speech

Voices that read text aloud.

Elo from blind human preference votes. Source

Frontier (API)

  1. Eleven v4ElevenLabs1339
  2. Eleven v4 TurboElevenLabs1335
  3. Qwen-Audio-3.1-TTS-PlusAlibaba1303
  4. Sonic 3.6Cartesia1297
  5. Simba 3.5SpeechifyAI1286
  6. Gemini 3.8 Flash TTSGoogle1283
  7. Lightning V3.1 Pro (Oct 2026)Smallest.ai1273
  8. Qwen-Audio-3.0-TTS-PlusAlibaba1268

Open weights (run it yourself)

  1. Fish Audio Drama 3Fish Audio1275
  2. Fish Audio Drama 3 PreviewFish Audio1241
  3. Breeze TTS 2BreezeBlue1228
  4. Ming-3.0-ttsAnt Group1187
  5. Phonon 31M (English)Gradium · 31M params1114
  6. Fish Audio S2 ProFish Audio1108
  7. Step Audio EditX (Mar 2026)StepFun1101
  8. Voxtral TTSMistral1092

Bars share one scale starting at 1029.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
Maya 2 NativeMaya Research · 2026-10-26Streaming text-to-speech for 11 Indic languages plus English, with 20 built-in voices and cross-lingual voice cloning from a 3–10 second clip.Ranked #2 among real-time models on the VoiceArena Hindi TTS leaderboard (Elo 1047 ±11, checked Oct 27, 2026), behind Cartesia Sonic-3.6 and ahead of Sarvam Bulbul-V3 and Grok TTS. The authors' own bench6k test (6,000 held-out texts) reports 3.40% overall CER using Qwen3-ASR as the judge.2.5Bfits?apache-2.0
Ming-3.0-tts 0.5BinclusionAI · 2026-09-22A small multilingual voice-cloning TTS model that can also design voices from a text description and add laughs, sighs, breaths, emotion and Sichuanese/Cantonese dialect on command.Authors' own README: Seed-TTS zh WER 0.80% / en WER 1.58%; best Chinese CER (2.70) and OCER (3.20) among compared open models on NVBench single-tag paralinguistic control; best CER on WSC-Eval-TTS-easy (2.27%) and WSYue-TTS-eval-Base (6.15%, 98% dialect accuracy).1.5Bfitsfitsmit
IndexTTS-2.5Bilibili IndexTeam · 2026-08-10Zero-shot cloning with emotion kept separate from timbre and precise speed/duration control, which suits dubbing (zh/en/ja/es/ar).Author states ~0.8B GPT backbone and roughly 6 GB VRAM for inference (model card)800Mfits?bilibili Model Use License Agreement
Higgs TTS 3 4BBoson AI · 2026-06-04High-quality expressive multilingual TTS covering 100+ languages.Model card claims single-digit WER/CER on 102 languages, under 5 on 85 of them (Boson AI model card, self-reported)4.7BtightfitsBoson Higgs TTS 3 Research and Non-Commercial License (plus a Creator Use Grant)non-commercial
Supertonic 3Supertone · 2026-05-06Tiny on-device TTS in 31 languages that runs on CPU via ONNX Runtime with no GPU needed.Author reports ~99M params and WER/CER in a range competitive with much larger models like VoxCPM2 (Supertone model card, self-reported)99MfitsfitsOpenRAIL-M (weights), MIT (sample code)
VoxCPM2OpenBMB · 2026-04-03Tokenizer-free 2B TTS with 48 kHz output in 30 languages, voice cloning, and real-time streaming.Author reports ~8 GB VRAM and RTF ~0.30 on an RTX 4090 (~0.13 with Nano-vLLM); full Seed-TTS-eval/CV3-eval tables are on GitHub (OpenBMB model card)2.3Bfitsfitsapache-2.0
OmniVoicek2-fsa (Next-gen Kaldi) · 2026-03-30Zero-shot voice cloning and voice design in 600+ languages, and very fast.Model card claims 600+ languages and RTF as low as 0.025 (40x real time) (k2-fsa model card, self-reported)610MfitsfitsCode Apache-2.0; weights CC-BY-NC (per model card)non-commercial
Qwen3-TTS 1.7BAlibaba Qwen · 2026-01-21Strong all-round multilingual TTS with 3-second voice cloning; sibling checkpoints add preset voices (CustomVoice) and voice design from a text description.Seed-TTS test-en WER 1.24 / test-zh 0.77, lower than CosyVoice 3 (1.45/0.71 zh) and MiniMax-Speech (1.65) in the authors' table (Qwen model card, self-reported)1.9Bfitsfitsapache-2.0
VibeVoice 1.5BMicrosoft · 2025-08-25Long-form, multi-speaker conversational audio such as podcasts; the 0.5B Realtime sibling does streaming TTS.2.7Bfitsfitsmit
Chatterbox (Multilingual v3 / Turbo)Resemble AI · 2025-04-24Popular MIT-licensed voice-cloning TTS with emotion exaggeration control and 23 languages; the 350M Turbo variant targets low-latency voice agents.Turbo is described as a 350M-parameter model with paralinguistic tags like [laugh] (ResembleAI/chatterbox-turbo card); multilingual repo updated to V3 on 2026-06-10—fitsfitsmit

Speech to text

Transcription accuracy.

Accuracy = 100 − word error rate (AA WER index). Source

Frontier (API)

  1. Fun-Realtime-ASR-previewAlibaba · 1.7% word error rate98.3
  2. MAI-Transcribe-2Microsoft AI · 2.0% word error rate98.0
  3. Voxtral Small Transcribe 3Mistral · 2.1% word error rate97.9
  4. Scribe v2ElevenLabs · 2.2% word error rate97.8
  5. Voxtral Mini Transcribe 3Mistral · 2.2% word error rate97.8
  6. Grok Voice Transcribe 2.0SpaceXAI · 2.3% word error rate97.7
  7. MAI-Transcribe-1.5Microsoft AI · 2.4% word error rate97.6
  8. StepAudio 3 ASRStepFun · 2.5% word error rate97.5

Open weights (run it yourself)

  1. Voxtral SmallMistral · 2.8% word error rate97.2
  2. Nemotron 3.5 TranscribeNVIDIA · 3.4% word error rate96.6
  3. Voxtral Mini TranscribeMistral · 3.8% word error rate96.2
  4. Whisper Large v2OpenAI · 4.1% word error rate95.9
  5. Whisper Large v3OpenAI · 4.1% word error rate95.9
  6. Canary Qwen 2.5BNVIDIA · 2.5B params · 4.3% word error rate95.7
  7. Parakeet TDT 0.6B V3NVIDIA · 600M params · 4.5% word error rate95.5
  8. Cohere TranscribeCohere · 4.6% word error rate95.4

Bars share one scale starting at 93.7.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
WhistleCactus Compute · 2026-09-30A tiny 16.9 MB speech-to-text model that runs on a CPU in English, German, French, Spanish, Italian, Dutch and Polish, with word timestamps and keyword biasing, for phones, wearables and microcontrollers.The authors' own model card compares word error rates with Whisper and Moonshine (their published figures) on the full test splits, including LibriSpeech, TED-LIUM, AMI, FLEURS and MLS, at 2–4-bit precision. The numbers appear only in a chart, and the authors say no test audio appears in the training data.—fitsfitsapache-2.0
Parakeet Redux (1.58-bit)Moondream · 2026-09-18Ternary-weight Parakeet that fits in 178 MB and transcribes ~113x real time on 8 CPU cores, which makes it good for laptops.Open ASR Leaderboard English avg WER 6.55 (vs 6.26 for the original) and FLEURS-25 10.56 (beats 11.62) (Moondream model card, self-reported)150Mfitsfitscc-by-4.0
Confucius4-R2T2NetEase Youdao · 2026-09-10Streaming speech-to-text for Chinese and English that emits words as it hears them and never rewrites what it already printed, which suits live captions and anything feeding a downstream pipeline.Authors' own evaluation reports 2.13% WER on LibriSpeech clean and 4.88% on LibriSpeech other at a 160 ms streaming chunk, versus 3.86%/9.64% for X-ASR and 4.70%/9.86% for WhisperRT at comparable latency, with 200–600 ms average latency; Chinese CER 5.87% on Wenet-net.2Bfitsfitsother
MOSS-Transcribe-Diarize 0.9BOpenMOSS (Fudan) · 2026-05-19Small all-in-one model for meetings and podcasts: speaker-labelled, timestamped transcripts of up to 90 minutes in one pass, 50+ languages.Won first place in the 2nd MLC-SLM Challenge at INTERSPEECH 2026 (model card news)910Mfitsfitsapache-2.0
Nemotron 3.5 ASR Streaming 0.6BNVIDIA · 2026-05-15Low-latency streaming ASR in 40 language-locales with punctuation and chunk sizes from 80 ms, built for live captions and voice agents.640MfitsfitsOpenMDW-1.1 (permissive)
Granite Speech 4.1 2BIBM · 2026-04-16Compact enterprise-friendly ASR plus speech translation (en/fr/de/es/pt/ja) with keyword biasing and punctuation.2.3Bfitsfitsapache-2.0
Qwen3-ASR 1.7BAlibaba Qwen · 2026-01-28Accurate multilingual ASR with language ID across 52 languages and dialects, including singing, plus streaming and vLLM serving.LibriSpeech clean/other WER 1.63/3.38, Common Voice-en 7.39 vs Whisper-large-v3 9.90 (Qwen model card, self-reported)2.4Bfitsfitsapache-2.0
VibeVoice-ASRMicrosoft · 2026-01-21Transcribes up to 60 minutes in one pass with built-in speaker diarization, timestamps, and hotwords (50+ languages).Technical report arXiv 2601.18184 (cpWER/tcpWER charts in the model card)8.7Bnofitsmit
Voxtral Mini 4B RealtimeMistral AI · 2026-01-21Streaming, real-time version of Voxtral (different from Voxtral Mini Transcribe) for live transcription up to hours long.4.4Btightfitsapache-2.0

Speech to speech

Real-time voice agents you talk to.

Elo from blind human preference votes. Source

Frontier (API)

  1. Cartesia Managed AgentsCartesia1111
  2. Gemini 3.1 Flash Live MinimalGoogle1089
  3. Gemini 3.8 LiveGoogle1074
  4. Gemini 3.1 Flash Live HighGoogle1062
  5. Grok Voice 3.0SpaceXAI1059
  6. GPT-Live-1 (Sol, low)OpenAI1037
  7. GPT-Live-1 (Astra, medium)OpenAI1027
  8. Phonic 1 (High)Phonic1012

Open weights (run it yourself)

No open-weight voice agent is ranked in this arena yet. The open models below listen and talk natively and run on your own hardware.

Bars share one scale starting at 986.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
Realtime-VenusinclusionAI (Ant Group) · 2026-09-16Brand-new full-duplex audio-visual agent built on MiniCPM-o 4.5 that can interrupt politely, speak proactively, and hand tasks to tools asynchronously.Card points to the Realtime-Venus technical report (arxiv.org/pdf/2609.13814) for audio-understanding and full-duplex results; numbers aren't reproduced in the card text9Bno?apache-2.0 (also review upstream MiniCPM-o terms)
Raon-SpeechChat 9BKRAFTON · 2026-04-01Full-duplex English voice chat with backchannels, barge-in, personas, and optional voice cloning.9.7Bno?cc-by-nc-4.0non-commercial
MiniCPM-o 4.5OpenBMB · 2026-02-039B omni model (vision + speech in and out) with full-duplex, proactive live streaming. It runs locally through llama.cpp/Ollama.Authors claim an OpenCompass average of 77.6, surpassing GPT-4o and Gemini 2.0 Pro. That is a vision benchmark, not a speech one (model card, self-reported)9.4Bfits quantizedfitsapache-2.0
PersonaPlex 7B v1NVIDIA · 2026-01-15Full-duplex English voice agent (Moshi-based) that listens while it talks, handles interruptions, and follows a text persona and voice prompt.FullDuplexBench: smooth turn-taking 0.908 success at 0.170 s latency, user interruption 0.950 at 0.240 s (NVIDIA model card, self-reported)8.4Bfits quantizedfits quantizedNVIDIA Open Model License
LFM2.5-Audio 1.5BLiquid AI · 2025-12-18Tiny end-to-end speech-to-speech chat model with interleaved text and audio output for low latency; it also does ASR and TTS.VoiceBench overall 54.92 vs Moshi 29.51 and Qwen2.5-Omni-3B 63.57 (LiquidAI model card, self-reported)1.5BfitsfitsLFM Open License v1.0 (commercial use free below $10M annual revenue)
Qwen3-Omni 30B-A3B InstructAlibaba Qwen · 2025-09-20The smartest open omni model that hears and speaks natively (Thinker-Talker MoE with 3B active params). It is turn-based rather than full-duplex.Model card lists 78.85 GB minimum GPU memory at BF16 for a 15 s video (Qwen model card)35.3Bnofits quantizedapache-2.0
Moshi (Moshiko/Moshika)Kyutai · 2024-09-11The original open full-duplex speech model, still the easiest local baseline, with official MLX and Rust/Candle quantized builds.VoiceBench overall 29.51 for Moshi vs 54.92 for LFM2.5-Audio, per LiquidAI's comparison table (LiquidAI model card)7.7Bfits quantizedfitscc-by-4.0

Audio to audio

Voice conversion, speech cleanup, stem splitting and sound isolation.

Notable open models

ModelGood atSize12GB GPU36GB MacLicense
Applio (RVC)IAHispano · 2026-09-19The actively maintained RVC voice-conversion suite for training your own voice models and converting speech or singing.Release 3.6.5 on 2026-09-19; ~3.7k GitHub stars (GitHub API)—fitsfitsmit
AuKTencent Hunyuan · 2026-09-09One instruction-driven model for speech editing (words, emotion, pitch, accent, timbre), enhancement, and speaker/vocal separation; it also does TTS.Benchmark chart covers generation, editing, enhancement, and separation (model card figure, arXiv 2609.08936)1.5Btight?mit
NVIDIA NemotronLabs RE-USENVIDIA · 2026-07-22A tiny 9.6M-parameter model that cleans up speech recordings in any language, fixing noise, echo, clipping, codec artifacts, dropped audio and low-quality mics at 8–48 kHz.On the URGENT 2025 non-blind test set (full set), it scores PESQ 2.57, STOI 0.78, speaker similarity 0.86 and ASR character accuracy 89.17. UniPASE (500M) scores 2.17, 0.76, 0.81 and 87.28, and Sidon (250M) scores 1.73, 0.62, 0.74 and 84.68 (authors' own model card; arXiv 2603.02641).10Mfitsnoopenmdw-1.1
HushWeya AI · 2026-03-19Real-time CPU speech enhancement for voice agents that also suppresses background speakers, not just noise.Author: 8 MB model, under 1 ms processing per 10 ms of audio on CPU, trained on 10,000+ hours (model card)2Mfitsfitsapache-2.0
LavaSR v2YatharthS (community) · 2026-02-1150 MB bandwidth-extension and denoise model that makes muffled or low-sample-rate speech sound full-band, extremely fast.Author claims ~5000 s of audio enhanced per second and quality above ~6 GB diffusion models (model card, self-reported)—fitsfitsapache-2.0
SAM AudioMeta · 2025-12-12Pulls any sound out of a mix from a text prompt, a video click, or a time span: 'isolate the dog barking' or 'just the vocals'.Official judge-model scores are listed per size in the GitHub README (facebookresearch/sam-audio)—?fitsSAM License (custom Meta; royalty-free use, includes acceptable-use/trade-control terms)
Seed-VC (v2)Plachtaa (community) · 2025-04-07Zero-shot voice and singing conversion from a 1-30 s reference, with a ~300 ms real-time mode; still the go-to despite the repo being archived.Author reports ~300 ms algorithm delay for real-time conversion, tested on an RTX 3060 Laptop (GitHub README)157Mfitsfitsgpl-3.0
Mel-Band RoFormer (vocals)Kimberley Jensen / community · 2024-08-06The go-to vocal/instrumental stem splitter family, usually run through python-audio-separator, UVR, or ZFTurbo's MSST.—fitsfitsmit
Demucs v4 (HTDemucs)Meta (now maintained by Alexandre Défossez)Reliable 4-stem split (vocals/drums/bass/other) that runs anywhere.—fitsfitsmit

Music generation

Full songs with vocals from a prompt.

Elo from blind human preference votes. Source

Frontier (API)

  1. Suno v6Suno1144
  2. Suno v6-miniSuno1132
  3. Mureka V9.5Mureka1101
  4. Mureka V9Mureka1063
  5. Lyria 3.5Google1050
  6. Lyria 3 ProGoogle1048
  7. Eleven Music v2.5ElevenLabs1016
  8. Eleven Music v2ElevenLabs1002

Open weights (run it yourself)

No open-weight model is ranked on the vocals board yet. YuE2, ACE-Step and the other open song generators below run on your own machine.

Bars share one scale starting at 966.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
MuLaCoverHeartMuLa (MuLa Labs) · 2026-09-14Generates cover versions and remixes of an existing song, steered by your own lyrics and a structured style prompt for genre, instruments and mood.4.4Btightfitscc-by-nc-4.0non-commercial
YuE2 3BM-A-P / HKUST · 2026-09-09Full songs with vocals from lyrics plus a style prompt, with an editable melody/chord score, covers, and agent-driven edits.WildSongBench (best-of-8) SongBench avg 6.9632 vs Suno v5 6.8721 and Suno v6 6.5562 (authors' benchmark, m-a-p/WildSongBench)3.6Btightfits quantizedcc-by-nc-4.0non-commercial
MiniMax Music 3MiniMax · 2026-08-07Five-minute full songs with expressive vocals and long-range structure, from a big-lab model.Author architecture: 8B global LLM + 0.6B local LLM + 2.4B flow-matching + 123M VAE decoder, 32 kHz stereo (model card)11.1Btight?MiniMax-Music3 Community License
Magenta RealTime 2Google DeepMind · 2026-05-28Live, streaming instrumental music you can steer in real time, for jamming and performance rather than finished songs.Author: 2.4B 'base' and 230M 'small' configurations built for on-device streaming (model card)2.4BfitsfitsWeights CC-BY-4.0, code Apache-2.0
Stable Audio 3 MediumStability AI · 2026-05-17Licensed-data music and sound generation up to ~6 minutes; Small Music (0.57B) and Small SFX variants run on-device.Up to 6:20 of audio and a 1.4B DiT per press coverage (the-decoder, MarkTechPost); 3.0 Large stays API-only (stability.ai post)2.3BfitsfitsStability AI Community License (free below $1M annual revenue)
SongGeneration 2 (LeVo 2) largeTencent AI Lab · 2026-03-01Hybrid LLM plus diffusion full-song generator (up to 4.5 min) in several languages.Authors report PER 8.55% and a 20-expert human evaluation (per vllm-omni issue #3390 summarizing official specs)4Bnofits quantizedUnclear: HF says 'unknown', GitHub says NOASSERTION (custom Tencent terms)
HeartMuLa oss 3BHeartMuLa team · 2026-02-13Apache-licensed lyrics-to-song model with good lyric control; the team also ships a codec and a lyric transcriber.Authors call the happy-new-year build the best open model for lyric controllability and music quality (heartlib README, self-claim)3.9Btight?apache-2.0
ACE-Step 1.5 (and 1.5 XL)ACE Studio + StepFun · 2026-01-23The practical, commercially safe local song generator: fast, trained on licensed data, with covers, repainting, LoRA training, and 50+ languages.Author claims a full song in under 10 s on an RTX 3090 and under 4 GB VRAM (model card); XL 4B DiT released 2026-04-022Bfitsfitsmit
DiffRhythm 2ASLP-lab / Xiaomi · 2025-10-13Fast diffusion-based full-song generation with vocals under a permissive license.1.1Bfits?apache-2.0

Image generation

Text to image.

Elo from blind human preference votes. Source

Frontier (API)

  1. GPT Image 2.6 Sunburst (max)OpenAI1259
  2. GPT Image 2.6 Flare (max)OpenAI1228
  3. GPT Image 2.5 Sunburst (max)OpenAI1227
  4. GPT Image 2.5 Flare (max)OpenAI1208
  5. GPT Image 2 (high)OpenAI1188
  6. Nano Banana 2.1Google1176
  7. MAI-Image-2.6Microsoft AI1168
  8. Grok Imagine Image 2.0SpaceXAI1166

Open weights (run it yourself)

  1. Qwen-Image-2.1Alibaba1045
  2. FLUX 3 Image KleinBlack Forest Labs1043
  3. Fibo 2.0Bria1039
  4. Ideogram 4.0 (Quality)Ideogram1027
  5. Ideogram 4.0Ideogram1003
  6. Ming-Image-0.1-DesignInclusionAI1000
  7. FLUX.2 [dev]Black Forest Labs1000
  8. FLUX.2 [dev] TurboFal996

Bars share one scale starting at 929.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
LLaDA-Image-TurboinclusionAI (Ant Group) · 2026-08-28New fully open-recipe 6B unified generate-and-edit model with a 4-step distilled variant; a permissive alternative to Qwen-Image-2.1.Author: state-of-the-art overall on Qwen-Image-Bench (53.53 EN / 53.38 ZH) for the base LLaDA-Image.6.5Bfits quantized?Apache-2.0
Krea 2 TurboKrea · 2026-06-22Krea's 12.8B from-scratch model, distilled to 8 steps for fast native 2K photoreal and aesthetic images; the Raw checkpoint (krea/Krea-2-Raw) is meant for fine-tuning.Author: ~2s native 2K images on consumer hardware (VentureBeat, krea.ai). No third-party arena number found.12.8Bfits quantizedfits quantizedKrea 2 Community License (free commercial use only under $1M annual revenue AND under 50 seats; content filtering required)
HiDream-O1-ImageHiDream.ai · 2026-05-08An 8B pixel-level unified transformer (no VAE, no separate text encoder) that does text-to-image up to 2048px, long-text rendering, editing and subject personalization.Card: HiDream-O1-Image-Dev-2604 (2026-05-13) debuted at #8 in the Artificial Analysis Text to Image Arena as the leading open-weights model at the time. Card also reports GenEval, DPG-Bench, HPSv3, CVTG-2K and LongText-Bench tables. It may already be on AA, so check before adding.8.8Bfits quantizedfitsMIT
FLUX.2 [klein] 4BBlack Forest Labs · 2026-01-14BFL's smallest FLUX.2 model: generation and multi-reference editing in about 4 steps, under Apache-2.0 (unlike the 9B klein and [dev]).Author: sub-second end-to-end inference, runs on consumer GPUs (~13GB VRAM), such as an RTX 3090/4070.3.9Bfits quantizedfitsApache-2.0
GLM-ImageZ.ai (Zhipu) · 2026-01-08Hybrid autoregressive (9B, GLM-4 based) plus 7B diffusion decoder that is especially good at dense text, posters and knowledge-heavy images, and also does image editing.Author: strong text rendering and knowledge-intensive generation (z.ai/blog/glm-image). The card documents ~23GB GPU memory with CPU offload.16BtighttightMIT
Z-Image-TurboAlibaba Tongyi-MAI · 2025-11-25Small 6B single-stream DiT that makes strong photoreal images and bilingual text in 8 steps; the most-downloaded new local image model of the past year.Author: state-of-the-art among open-source models on the Elo-based Alibaba AI Arena (aiarena.alibaba-inc.com). Card says it fits in 16GB VRAM.6.2Bfits quantizedfitsApache-2.0

Image editing

Change an existing image with instructions.

Elo from blind human preference votes. Source

Frontier (API)

  1. GPT Image 2.6 Sunburst (max)OpenAI1209
  2. GPT Image 2.6 Flare (max)OpenAI1206
  3. GPT Image 2.5 Sunburst (max)OpenAI1199
  4. MAI-Image-2.7-PreviewMicrosoft AI1195
  5. GPT Image 2.5 Flare (max)OpenAI1182
  6. Krea 3 (Ultra)Krea1152
  7. Nano Banana 2.1Google1152
  8. MAI-Image-2.6Microsoft AI1149

Open weights (run it yourself)

  1. Fibo 2.0Bria1085
  2. Qwen-Image-2.1Alibaba1080
  3. FLUX 3 Image KleinBlack Forest Labs1069
  4. HunyuanImage 3.0 InstructTencent1065
  5. FLUX.2 [dev] TurboFal1017
  6. FLUX.2 [dev] FlashFal1013
  7. FLUX.2 [klein] 9BBlack Forest Labs · 9B params1013
  8. FLUX.2 [dev]Black Forest Labs1000

Bars share one scale starting at 947.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
SenseNova-U1.5-8B-MoTSenseNova (SenseTime) · 2026-08-19A single unified model that both generates images and edits them from plain instructions, with particular attention to keeping faces, backgrounds and untouched regions intact and to rendering legible Chinese and English text in posters and infographics.No numeric results appear in the card text; the benchmark section is chart images only. The authors' own claims are improved native 4K generation and more reliable local, text, multi-reference and insertion edits, described in their NEO-unify blog post and the SenseNova-U1.5 paper (arXiv 2609.11929).17.5Bnonoapache-2.0
JoyAI-Image Edit PlusJD Open Source · 2026-06-22A 16B MMDiT multi-image editor with a Qwen3-VL-8B encoder, and the newest large permissive editor after Qwen's.16.3Btightfits quantizedApache-2.0
HiDream-O1-ImageHiDream.ai · 2026-05-08A unified 8B pixel-space model whose full (non-Dev) checkpoint is recommended for instruction editing and subject-driven personalization.Card: HiDream recommends the full model for editing tasks. No edit-benchmark number was extracted.8.8Bfits quantizedfitsMIT
FireRed-Image-Edit-1.1FireRed Team (Xiaohongshu) · 2026-03-02A 20B Qwen-Image-based editor tuned for identity consistency and prompt following, under a permissive license.Author: new open-source state of the art on ImgEdit, GEdit and RedEdit, and open-source SOTA in identity preservation. The acceleration suite gives 4.5s end-to-end at 30GB VRAM.20.4Btightfits quantizedApache-2.0
LongCat-Image-Edit-TurboMeituan LongCat · 2026-02-03Distilled fast version of Meituan's bilingual editor. The DiT is about half Qwen-Image's size, and quality is strong for multi-turn edits.Author: LongCat-Image-Edit achieves state-of-the-art among open-source image editing models (card claim, no table in the card).—fits quantizedfits quantizedApache-2.0
FLUX.2 [klein] 4BBlack Forest Labs · 2026-01-14The Apache-licensed 4B sibling of the klein 9B already on the page; it does multi-reference editing in a few steps and is the easiest local editor to run.Author: runs on consumer GPUs (~13GB VRAM), sub-second inference.3.9Bfits quantizedfitsApache-2.0

Video generation

Text to video.

Elo from blind human preference votes. Source

Frontier (API)

  1. Wan 3.0Alibaba1172
  2. Utopai XUtopai Studios1167
  3. FLUX 3 MaxFal1156
  4. Dreamina Seedance 2.5ByteDance Seed1155
  5. HeyGen Video 1HeyGen1152
  6. MiniMax H3 MaxFal1145
  7. FLUX 3Black Forest Labs1138
  8. Dreamina Seedance 2.0ByteDance Seed1135

Open weights (run it yourself)

  1. MiniMax H3 (768p)MiniMax1160
  2. LTX-2.5 FastLightricks942
  3. LTX-2.5 ProLightricks941
  4. LTX-2.3 ProLightricks883
  5. LTX-2.3 FastLightricks838

Bars share one scale starting at 754.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
FastH3 4-step (MiniMax H3 distilled)FastVideo (Hao AI Lab) on MiniMax H3 · 2026-08-27A 4-step distillation of MiniMax H3 (already on the page) for text-to-audio+video, and the most-downloaded way to run H3 fast locally.Author: text-to-audio-video with four transformer forwards; preview quality below the base on hard motion. No third-party benchmark.35BnotightMiniMax H3 Community License (inherits base-model terms; commercial terms not verified)
Wan-Animate-2 (Wan2.2-Animate-2-14B)Alibaba Wan-AI · 2026-08-07Wan's newest open release: end-to-end character animation and replacement driven by a reference video, with base and distilled weights.14Bno?Apache-2.0
JoyAI-Video-EditJD.com (jd-opensource) · 2026-08-04Edits a live or uploaded video from a text instruction frame by frame as it streams, handling subject swaps, background and style changes, and reference-image-guided edits without needing the whole clip first.Authors report the full pipeline runs at 30 FPS at 720×1248, and at 840×480 / 24 FPS on a single 32GB RTX 5090 (model card and arXiv 2608.03974).16Bnonoapache-2.0
LTX-2.5Lightricks · 2026-07-23Image-to-video generation from the LTX family, known for fast generation of short clips on consumer hardware.—??other
LongCat-VideoMeituan LongCat · 2025-10-24A 13.6B dense model for text-to-video, image-to-video and video continuation, built for minutes-long 720p/30fps videos without color drift.Author: internal MOS benchmark comparable to leading open and closed models (vs Wan 2.2 A14B, 28B total).13.6BtighttightMIT
Krea Realtime Video 14BKrea · 2025-10-08An autoregressive, self-forcing distillation of Wan 2.1 14B that streams text-to-video in real time and lets you steer prompts mid-generation.14.3Btight?Apache-2.0
Kandinsky 5.0 Video LiteKandinsky Lab (Sber AI) · 2025-09-29A 2B text-to-video model (5s/10s, with 16-step distilled variants) that is among the lightest capable open video models.Author: ranks #1 among open-source models in its class and outperforms Wan 5B and 14B in side-by-side on Movie Gen prompts. VBench chart in the card.2Bfits quantizedfitsMIT
Wan2.2-TI2V-5BAlibaba Wan-AI · 2025-07-28The small dense Wan 2.2 model: text-to-video and image-to-video at 720p/24fps with a high-compression VAE. It's the practical local Wan (the page only lists the 14B ones).Author: one of the fastest 720p@24fps models, runs on a single consumer GPU such as an RTX 4090.5Bfits quantizedfitsApache-2.0

Upscaling and restoration

Sharpen, enlarge and repair images and video.

Notable open models

ModelGood atSize12GB GPU36GB MacLicense
RealRestorerRealRestorer team · 2026-03-25A general real-world photo restoration model (deblur, denoise, compression, low-light and more) fine-tuned from the Step1X-Edit editor, with its own RealIR-Bench.Paper arXiv 2603.25502; RealIR-Bench dataset released.12.4BnotightApache-2.0
FlashVSR v1.1Junhao Zhuang et al. (Tsinghua / Shanghai AI Lab) · 2025-10-31One-step streaming diffusion video super-resolution built on Wan 2.1, aimed at near-real-time 4x video upscaling.Author: ~17 FPS for 768x1408 video on a single A100.—??Apache-2.0
SeedVR2-7BByteDance Seed · 2025-06-11Larger SeedVR2 with a 'sharp' variant for maximum detail.Paper arXiv 2506.05301.7Btightfits quantizedApache-2.0
SeedVR2-3BByteDance Seed · 2025-06-10One-step diffusion restoration and upscaling for both images and video, and the de facto current go-to upscaler in ComfyUI.Paper arXiv 2506.05301. Adoption: numz/SeedVR2_comfyUI (471k downloads) and Comfy-Org/SeedVR2 (441k downloads) on HF.3Bfits quantizedfitsApache-2.0
4x-UltraSharp V2Kim2091 · 2025-05-12A small, fast GAN-style 4x image upscaler that is the community default final pass in ComfyUI/A1111 workflows.—fitsfitsCC BY-NC-SA 4.0 (non-commercial)non-commercial
Real-ESRGANXintao Wang (Tencent ARC) · 2022-09-20The classic permissive, lightweight 4x image and anime upscaler. It's older, but still the safe commercial-OK fallback.GitHub: 36.9k stars. Latest release v0.3.0 (2022-09-20). Last push 2024-08-06.—fitsfitsBSD-3-Clause

3D generation

Textured 3D models from an image or a prompt (most text-to-3D tools render an image first).

Elo from 3D Arena render votes (image to 3D). Source

Frontier (API)

  1. PicGen3D1415
  2. Meshy-61350
  3. 404_GEN1334
  4. Meshy-71311
  5. Zaohaowu3D1294
  6. MeshFormer1171
  7. Meshy-51001

Open weights (run it yourself)

  1. TRELLIS-3DGS1397
  2. TRELLIS1303
  3. Hunyuan3D-21283
  4. Hunyuan3D-2.11272
  5. InstantMesh1269
  6. SAM-3D-Objects-3DGS1267
  7. UniLat3D-3DGS1262
  8. Pixal3D1260

Bars share one scale starting at 897.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
TripoSplatVAST / Tripo AI · 2026-05-31Turns one image straight into 3D Gaussian splats (.ply/.splat) with a learned, variable number of Gaussians, and ships with an official ComfyUI template.Input: image—??MIT
HY-World 2.0Tencent Hunyuan · 2026-04-16Scene-scale 3D: text or an image becomes an explorable 3DGS or mesh world that exports to Unity or Unreal, and WorldMirror 2.0 rebuilds 3D from photos or video.Input: text+imageREADME reports WorldMirror point-map accuracy and completeness on 7-Scenes, NRGBD and DTU against Pow3R and MapAnything (github.com/Tencent-Hunyuan/HY-World-2.0).1.2B?noTencent HY-World 2.0 Community License (not valid in the EU, UK or South Korea; a separate license is needed above 1M MAU)
AniGenVAST / Tripo AI · 2026-03-03Generates an animation-ready asset (mesh plus skeleton and skinning) from a single image, built on TRELLIS, so it covers rigging as well as shape.Input: image—nonoMIT
TRELLIS.2-4BMicrosoft · 2025-12-01The newer version of TRELLIS: a single image becomes a high-res mesh with full PBR materials, including open surfaces and translucency, under a clean MIT license.Input: imageModel card calls it state-of-the-art; generates at 512^3 to 1536^3 voxel resolution (github.com/microsoft/TRELLIS.2). Most downloaded image-to-3D model on HF (1.77M).4Bno?MIT
Hunyuan3D-OmniTencent Hunyuan · 2025-09-25An image-to-shape model built on Hunyuan3D 2.1 that also accepts point clouds, voxels, bounding boxes or a skeleton pose to control the geometry.Input: image3.3Bfits?Tencent Hunyuan 3D Omni Community License (not valid in the EU, UK or South Korea; a separate license is needed above 1M MAU)
Cube 3D v0.5 (+ CubePart)Roblox · 2025-07-20A native text-to-shape model, rare among open 3D releases, and one of the few whose README documents Apple Silicon support; its May 2026 CubePart add-on (Roblox/cubepart) splits a mesh into named, game-ready parts.Input: text—tightfitsOpenRAIL
Step1X-3DStepFun · 2025-05-13A fully open (weights, training code and an 800K-asset curated data list) two-stage image-to-3D pipeline that outputs a watertight mesh plus texture, under Apache 2.0.Input: imageREADME benchmark: geometry plus texture uses 27GB of GPU memory and takes 152s for 50 steps.4.8Bno?Apache-2.0
TripoSGVAST / Tripo AI · 2025-03-24A 1.5B rectified-flow image-to-shape model with sharp, clean geometry that runs on modest consumer GPUs; the scribble variant (TripoSG-scribble) adds prompt-guided sketch input.Input: image1.4Bfits?MIT

World models

Models that learn how the world moves: JEPA-style video understanding and interactive worlds you can walk through.

Notable open models

ModelGood atSize12GB GPU36GB MacLicense
Hy-NomasTencent Hunyuan · 2026-10-21Generates a 720p, 24 FPS video world one segment at a time, and you can steer each segment with text instructions (add objects, change style or scene, describe actions) and camera paths.No benchmark results on the card. The authors released their own Hy-Nomas-Bench (280 test cases covering multi-turn text interaction, camera navigation and spatial memory), and they claim an 8-step distilled DiT runs on a single RTX 4090 with INT8 weights and CPU offload (authors' own claim).—nonoapache-2.0
LingBot-World 2.0 (Infinity)Robbyant (Ant Group) · 2026-07-09An open world simulator with minute-scale memory, conditioned on camera poses or actions plus text, released in 14B and 1.3B sizes with distilled real-time variants.Input: actions1.7B?noCC-BY-NC-SA-4.0 (v1, robbyant/lingbot-world-base-cam, was Apache-2.0)non-commercial
ABot-World-0AMAP CV Lab (Alibaba) · 2026-07-09The most desktop-friendly open world model found: infinite action-driven 720p rollouts on a single consumer GPU, with its 500-hour action-labeled training set also released.Input: actionsAuthors: 720p, 16 FPS, 1.2s latency and 19GB of GPU memory on one RTX 5090 (github.com/amap-cvlab/ABot-World).5.3BnonoApache-2.0
Qwen-AgentWorld-35B-A3BQwen · 2026-06-22Predicts what an environment would do next — terminal output, tool responses, web and Android screens — so agents can be trained and tested without running the real thing.Authors' own AgentWorldBench scores it 56.39 overall across seven domains, versus 47.73 for its Qwen3.5-35B-A3B base model and 56.59 for Claude Opus 4.8 (model card).34.7Bnofits quantizedapache-2.0
Matrix-Game 3.0Skywork AI · 2026-03-27An Apache-licensed interactive world model that streams 720p video in real time from an image and text prompt plus keyboard and mouse actions, with long-horizon memory.Input: actionsAuthors report 720p at 40 FPS with the 5B distilled model plus INT8 (arXiv 2604.08995). A 28B MoE version was announced but not yet released.5BnonoApache-2.0
Cosmos3-NanoNVIDIA · 2026-03-10NVIDIA's physical-AI world foundation model, now one omni model that takes text, image, video and action input and generates video, images and actions for robotics and driving simulation.Input: text+image15.8BnonoOpenMDW-1.1
HY-World 1.5 (WorldPlay)Tencent Hunyuan · 2025-12-17A real-time, playable world model from an image or text prompt, driven by keyboard and mouse, with geometric memory so places look the same when you come back; the full training stack is open too.Input: actions8BnonoTencent HY-WorldPlay Community License (not valid in the EU, UK or South Korea; a separate license is needed above 1M MAU)
V-JEPA 2 / V-JEPA 2.1 (+ V-JEPA 2-AC)Meta FAIR · 2025-06-25Yann LeCun's JEPA line: self-supervised video models that learn how the physical world moves without generating pixels; the action-conditioned V-JEPA 2-AC plans robot actions, and V-JEPA 2.1 (Mar 2026) adds much better dense features.Input: videoREADME: SSv2 probe 77.3% vs 69.7% (InternVideo2-1B), Diving48 90.2%, EK100 anticipation 39.7% vs 27.6% (github.com/facebookresearch/vjepa2).1BfitsfitsMIT (most code and weights; some parts Apache-2.0; this ViT-g HF repo is tagged Apache-2.0)

OCR and document parsing

PDFs and scans to clean text, tables and formulas.

OmniDocBench v1.6 overall. Source

Frontier (API)

  1. Gemini 3 ProGoogle92.9
  2. Gemini 3 FlashGoogle92.6
  3. GPT-5.2OpenAI86.6

Open weights (run it yourself)

  1. TeleOCR (formerly NaviDC-OCR)StarDoc-AI / TeleAI · 1.4B params96.9
  2. OvisOCR2Alibaba (ATH-MaaS / Ovis) · 853M params96.5
  3. PaddleOCR-VL 1.6Baidu PaddlePaddle · 959M params96.3
  4. MinerU2.5-Pro (2605)OpenDataLab (Shanghai AI Lab) · 1.2B params95.8
  5. GLM-OCRZ.ai (Zhipu) · 1.3B params95.2
  6. Unlimited-OCRBaidu · 3.3B params94.0

Bars share one scale starting at 83.0.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
MiniCPM-OCR 1.0OpenBMB · 2026-10-21A small 1.3B model that turns Chinese and English documents, including scans, photos and curved pages, into Markdown with HTML tables and LaTeX formulas, and can also pull named fields out as JSON.The authors report 93.68 on Real5-OmniDocBench, 54.7 on OCRBench v2, and 1.2 pages/s on an RTX 5090. They say that is about 1.49× the throughput of PaddleOCR-VL-1.6 and 3.62× that of MinerU2.5 Pro (source: model card).1.3Bfitsfitsapache-2.0
Indra OCRDatalab · 2026-10-10A tiny CPU-only OCR and layout model that turns clean printed PDFs and scans into markdown or JSON with boxes and reading order, far faster than Tesseract or PaddleOCR.Authors' own results: 62.1 headline score on OmniParseBench, Datalab's own benchmark (66.8 with --math), against 58.4 for PaddleOCR PP-StructureV3. Median 228 ms per page on 4 x86 vCPUs, and 49.4 pages/s on 32 cores against 2.92 for Tesseract 5.1Mfitsfitsopenrail
Youtu-Parsing-OmniTencent Youtu Lab · 2026-10-09A 5B parser that turns document pages, charts, flowcharts, geometry figures, audio clips and videos into structured JSON with layout boxes, tables, formulas, transcripts and timestamps.96.96 Overall on OmniDocBench v1.6 (TeleOCR 96.91, PaddleOCR-VL-1.6 96.34). Also the best open-weight model on OmniParsingBench at 75.08 average, behind only Gemini-3-Pro. Both results are the authors' own, from the model card.5.3Btight?other
Xiaomi-OCR-0Xiaomi (SeerRay Lab) · 2026-09-29A 0.8B model that turns scanned or photographed documents into Markdown, with tables in OTSL and formulas in LaTeX, and that also answers questions about documents and extracts fields into a JSON schema you give it.Authors' own model card / tech report (arXiv 2609.36136): #1 on Real5-OmniDocBench (95.24) and 96.83 on OmniDocBench v1.6, essentially level with TeleOCR (96.87, 1.2B) and ahead of OvisOCR2 (96.58), PaddleOCR-VL-1.6 (96.33) and GLM-OCR (95.22). It averages 83.2 across five OCR-VQA benchmarks, beating Qwen3.5-2B (80.9).873Mfitsfitsapache-2.0
Jina-OCR-v1Jina AI · 2026-09-01Converts document pages and PDFs to clean Markdown with tables and LaTeX math, and does it fast on a small GPU because only about 570M of its 3B parameters are active per token.Authors report 91.14 overall on OmniDocBench v1.6 and 83.4 on olmOCR-Bench (+7.4 over the DeepSeek-OCR backbone), at 2.57 pages/s on a single A100 at concurrency 32 — the highest page throughput of the fourteen systems they measured (model card and technical report, arXiv:2609.03181).3.4Bfitsfitscc-by-nc-4.0non-commercial
Surya OCR 2Datalab (datalab-to) · 2026-05-14A 650M-parameter document model that handles layout, reading order, full-page OCR, and table structure in one pass, small enough to run on modest hardware.Authors report 83.3 on olmOCR-bench at 0.65B params — best in class under 3B on their chart, with comparison scores taken from the olmOCR-bench dataset card — plus 87.2% average on their internal 91-language benchmark and 5.35 pages/s on an RTX 5090 via vLLM.686Mfitsfitsopenrailnon-commercial
dots.mocrrednote (dots-studio) · 2026-03-19Multilingual successor to dots.ocr, one model for layout, text, tables and formulas.olmOCR-Bench 83.9 ± 0.9 (self-reported on card). Predecessor dots.ocr: OmniDocBench v1.6 90.77.3Bfitsfitsmit
Chandra OCR 2Datalab · 2026-03-16Highest olmOCR-Bench score of any open model here; handles handwriting, forms, tables and math.olmOCR-Bench 85.8 ± 0.8 (self-reported on card, SOTA claim). Not on the OmniDocBench leaderboard; Falcon-OCR card measured it at 88.06 on OmniDocBench.5.3BtightfitsModified OpenRAIL-M (free for research, personal use and startups under $2M funding/revenue; no use competing with Datalab's API)
LightOnOCR-2-1BLightOn · 2026-01-16Fast 1B end-to-end PDF-to-text model, Apache-licensed, with bounding-box variants for images.olmOCR-Bench 83.2 ± 0.9 (headers/footers excluded; LightOn HF blog, self-reported); 5.71 pages/s on one H100.1Bfitsfitsapache-2.0
DeepSeek-OCRDeepSeek AI · 2025-10-17Converts document pages and PDFs to markdown, with resolution presets that let you spend far fewer vision tokens per page when you need throughput.The authors' own paper, "DeepSeek-OCR: Contexts Optical Compression" (arXiv 2510.18234), reports roughly 97% OCR decoding precision when a page is compressed to under 10x fewer vision tokens than its text tokens; the model card itself lists no benchmark numbers.3.3Bfitsfitsmit

Text classification

Embedding models that sort text into labels.

MTEB(eng, v2) classification mean (%). Source

Frontier (API)

  1. Seed1.6-embeddingBytedance92.4
  2. ingot-8b-r3jcorners · 7.6B params90.4
  3. gemini-embedding-001google90.0
  4. Seed1.5-EmbeddingByteDance-Seed89.9
  5. text-embedding-004google86.0
  6. text-embedding-005google86.0
  7. text-multilingual-embedding-002google84.7
  8. Cohere-embed-english-v3.0Cohere79.7

Open weights (run it yourself)

  1. F2LLM-4Bcodefuse-ai · 4B params91.7
  2. F2LLM-v2-14Bcodefuse-ai · 14B params91.7
  3. F2LLM-v2-4Bcodefuse-ai · 4B params91.7
  4. Giga-Embeddings-instruct-10B-A1.8B-0826ai-sage · 10.5B params91.6
  5. F2LLM-v2-8Bcodefuse-ai · 7.6B params91.5
  6. Giga-Embeddings-instruct-3B-0826ai-sage · 3.2B params91.5
  7. F2LLM-v2-1.7Bcodefuse-ai · 1.7B params91.4
  8. F2LLM-1.7Bcodefuse-ai · 1.7B params90.9

Bars share one scale starting at 75.5.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
WeMM-Embedding-2BTencent (WeChat) · 2026-08-25A small embedding model that puts text, images, videos and document screenshots into one search space, for multimodal search and RAG.MMEB-v2 average 77.9 vs 73.2 for Qwen3-VL-Embedding-2B and 77.8 for Qwen3-VL-Embedding-8B (authors' own tech report, Table 1); 256-dim Matryoshka embeddings keep 98.7% of image/video performance (authors' claim).2.7Bfitsfitsother
GLiGuard 300MFastino · 2026-05-09Encoder guardrail model that checks prompt and response safety in one fast pass instead of calling a decoder guard LLM.Card: 87.7 avg F1 prompt harmfulness, 82.7 avg F1 response harmfulness over 9 benchmarks (self-reported).208Mfitsfitsapache-2.0
Granite Guardian 4.1 8BIBM · 2026-04-16Guard model for harm, jailbreak, RAG groundedness and custom criteria, also usable as a reward model.Card: OOD safety aggregate F1 0.79 (non-think); JETTS best-of-N 70.29 (self-reported).8.4Bfits quantizedfitsapache-2.0
GLiClass-instruct-large v1.0Knowledgator · 2026-02-12Zero-shot classifier that scores many labels in one pass, far faster than an LLM, and accepts instructions and few-shot examples.Card: average zero-shot F1 0.7199 across 14 text-classification datasets (self-reported).439Mfitsfitsapache-2.0
Qwen3Guard-Gen-4BQwen · 2025-09-23Safety classifier that labels prompts and responses as safe, controversial or unsafe with a category, in 119 languages (0.6B and 8B also available).Card claims state-of-the-art on English, Chinese and multilingual safety benchmarks (no single headline number on card).4.4Bfitsfitsapache-2.0
gpt-oss-safeguard-20bOpenAI · 2025-09-18Bring-your-own-policy safety classifier: you write the policy in plain text and it reasons about whether content breaks it.Card: 21B total / 3.6B active; OpenAI says it fits in 16GB of VRAM.21.5Btightfitsapache-2.0
GLiClass-edge v3.0Knowledgator · 2025-07-20A 33M-parameter zero-shot classifier for edge and CPU use, ~97 examples/s at batch size 1 on an A6000.Card: average benchmark F1 0.4873 vs 0.7001 for gliclass-large-v3.0 (self-reported); speed 97.29 ex/s on A6000.33Mfitsfitsapache-2.0
Llama Prompt Guard 2 86MMeta · 2025-04-28Small classifier that flags prompt injection and jailbreak attempts before they reach your LLM (a 22M version also exists).279Mfitsfitsllama4 (Llama 4 Community License)
ModernBERT-large-zeroshot-v2.0Moritz Laurer · 2024-12-27NLI-style zero-shot classifier on ModernBERT: pass any labels, get probabilities, with long context and bf16 speed.Card: accuracy per dataset across 28+ tasks; author notes it is slightly worse than deberta-v3-large-zeroshot-v2.0 on average but faster and more memory efficient.396Mfitsfitsapache-2.0
ModernBERT-baseAnswer.AI / LightOn · 2024-12-11The standard encoder to fine-tune your own fast classifier (8k context, 150M params).No leaderboard score. It is a base model you fine-tune yourself, and popular at ~4.3M HF downloads per month.150Mfitsfitsapache-2.0

Rerankers

Re-order search results for RAG and search.

MTEB(eng, v2) retrieval NDCG@10 after reranking (published by the ettin-reranker authors). Source

Frontier (API)

Hosted rerankers (Cohere, Voyage, Jina's API) aren't scored on this board.

Open weights (run it yourself)

  1. Qwen3-Reranker-4BQwen · 4B params63.7
  2. mxbai-rerank-large-v2Mixedbread · 1.5B params61.1
  3. ettin-reranker-1b-v1Sentence Transformers (Tom Aarsen) / JHU ettin · 1B params61.1
  4. ettin-reranker-150m-v1Sentence Transformers (Tom Aarsen) / JHU ettin · 149M params59.9
  5. Qwen3-Reranker-0.6BQwen · 596M params59.4
  6. gte-reranker-modernbert-baseAlibaba NLP · 150M params58.4
  7. bge-reranker-v2-m3BAAI · 568M params55.3
  8. zerank-2ZeroEntropy · 4B params53.0

Bars share one scale starting at 49.3.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
jina-reranker-v3.5Jina AI · 2026-07-14New 0.6B listwise reranker that matches or beats much larger rerankers on BEIR.Own card: BEIR 63.20 vs 62.28 for Qwen3-Reranker-4B and 62.45 for mxbai-rerank-large-v2 (self-reported). Not in the ettin comparison.597Mfitsfitscc-by-nc-4.0non-commercial
Qwen3-VL-Reranker-8BQwen · 2026-01-07Reranks search results where the query or documents can be any mix of text, images, screenshots or video, in 30+ languages, which makes it a good second stage for multimodal and visual-document RAG.Authors' own results (model card / arXiv 2601.04720): 79.2 average on MMEB-v2 retrieval, 83.6 on JinaVDR and 66.7 on ViDoRe v3, compared with 82.2 and 57.8 for jina-reranker-m0.8.8Bnofitsapache-2.0

Other open specialists

Small models that do one job well: next-edit code suggestions, tool calling, translation.

Notable open models

ModelGood atSize12GB GPU36GB MacLicense
PLaMo 3 Translate 31B (also 8B / 2B / 0.8B)Preferred Networks (PFN) · 2026-10-26A translation-only model for 53 languages, strongest in Japanese↔English, that can follow a glossary, use surrounding context and keep consistent with earlier translations.The authors' own model card reports the 31B at 0.8799 WMT24++ COMET-22 (En→Ja) and 56.61 FLORES+ chrF++ (En↔multilingual). The authors also claim that even their 2B model beats TranslateGemma and HY-MT2 on every metric they compared.31.8Bnofits quantizedother
Index-Translate-9BIndexTeam (Bilibili) · 2026-09-28Translates text across 150 languages while following glossaries, keeping JSON, markdown and placeholders intact, and adapting tone, slang and memes as instructed.Authors' technical report: instTrans IFscore 0.8209 vs 0.3068 for TranslateGemma-12B; FLORES COMET-22 0.8789; lowest low-resource off-target rate among the compared systems (3.47% on instTrans_minor).9.7Bnofitsapache-2.0
Needle 3Cactus Compute · 2026-09-16A tiny on-device model that turns user requests into the right function calls and typed JSON fields, and also produces embeddings, in a single file of a few megabytes.The authors say it beats models 10x its size on mobile tool calls and matches models 2-3x bigger on extraction. They also say that after fine-tuning on DroidCall, subnetworks from 4 layers (29M params) up pass DeepSeek V4 Flash. Both claims are from the model card and cactuscompute.com/needle.121Mfitsfitsapache-2.0
Hy-MT2 1.8BTencent Hunyuan · 2026-05-11A small, fast translation model covering 33 languages that follows instructions on terminology, style and structured-data formats, and it is light enough to run on a device.The authors' own model card and report (arXiv 2605.22064) say the 1.8B model beats commercial translation APIs from Microsoft and Doubao overall, and that the 7B and 30B-A3B sizes beat DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode.2Bfitsfitsapache-2.0
Zeta 2.1Zed Industries · 2026-05-07Open next-edit prediction model behind Zed's edit predictions, suggesting your next code change from recent edits.Card: fine-tuned from ByteDance-Seed/Seed-Coder-8B-Base; Apache-2.0. No benchmark on card.8.3Bfits quantizedfitsapache-2.0
Sweep Next-Edit 1.5BSweep · 2026-01-20Laptop-sized next-edit autocomplete (under 500ms locally) for JetBrains/VS Code-style tab-to-jump edits; 0.5B and 7B v2 also available.Card: 'outperforms models over 4x its size on next-edit benchmarks' (self-reported, see blog.sweep.dev/posts/oss-next-edit).1.5Bfitsfitsapache-2.0
TranslateGemma 12BGoogle · 2026-01-12Open translation model for 55 languages that beats the general Gemma 3 27B at translation (4B and 27B also available).TranslateGemma tech report (arXiv 2601.09012): WMT24++ MetricX 3.60 vs 4.04 for Gemma 3 27B (lower is better).13.2Bfits quantizedfits quantizedgemma (Gemma Terms of Use)
FunctionGemma 270MGoogle · 2025-10-08Tiny function-calling model meant to be fine-tuned into on-device agents that turn requests into API calls.Google: fine-tuning on the Mobile Actions task raised accuracy from 58% to 85% (blog.google FunctionGemma post).268Mfitsfitsgemma (Gemma Terms of Use)

System 1 models

Fast decisions, no text

Chat models think out loud: they generate an answer one word at a time, which is slow and can go off the rails. System 1 models are the fast, intuitive half. You hand one the situation (an email, a ticket, some JSON) and a set of typed questions, and it answers every question at once with a calibrated probability, in a fraction of a second. It never writes text, so there's nothing to parse and nothing to hallucinate. TypeSafe AI's Jev, released in September 2026, started the category, and open copies appeared within days.

The ranking is an independent benchmark that puts Jev against general-purpose LLMs forced to pick from the same options. Each row shows latency, cost per 1,000 decisions and calibration error (ECE, lower is better). No open System 1 model has been benchmarked independently yet.

System 1 decision models

Typed, calibrated decisions in one pass instead of generated text.

Accuracy on DMB's 77-way banking-intent suite (typed decisions). Source

Frontier (API)

  1. Jev (System One)TypeSafe AI · 274 ms · $0.07/1k · ECE 0.0876.3
  2. Claude Haiku 4.5Anthropic · 4360 ms · $0.82/1k · ECE 0.0875.7
  3. Claude Sonnet 4.6Anthropic · 2371 ms · $2.48/1k · ECE 0.0874.8
  4. GPT-5.4 miniOpenAI · 660 ms · $0.69/1k · ECE 0.2374.0
  5. GPT-5.4 nanoOpenAI · 684 ms · $0.19/1k · ECE 0.0970.9

Open weights (run it yourself)

  1. gpt-oss-120bOpenAI · 331 ms · $0.32/1k · ECE 0.0981.3
  2. GLM-5.3Z AI · 3425 ms · $2.42/1k · ECE 0.0480.4
  3. GLM-5.3 FlashZ AI · 3331 ms · $0.21/1k · ECE 0.0579.0
  4. DeepSeek ChatDeepSeek · 776 ms · $0.27/1k · ECE 0.0876.2

Bars share one scale starting at 67.3.

Also worth a look: open models the leaderboard hasn't ranked

ModelGood atSize12GB GPU36GB MacLicense
MiniCPM-OCR JevOpenBMB · 2026-10-23Scores how likely an OCR result for a text, formula or table region is to be correct (0 to 1), so you can auto-accept good regions, send doubtful ones to human review, or pick the best of several OCR engines.Decision AUROC (×100) of 93.6 on JevOCRBench, compared with 90.1 for JEV-35B-A3B and 93.8 for Qwen3.5-397B-A17B. Across four parsers (MiniCPM-OCR 1.0, GLM-OCR, PaddleOCR-VL-1.6, Xiaomi-OCR-0) on 200 samples it scores 91.9–93.4 (authors' own results, model card).2.2Bfitsfitsapache-2.0
elephant-1Base Labs (Baseten) · 2026-10-19Takes text or images plus a set of yes/no, multiple-choice or scored questions and returns calibrated probabilities for every answer in one pass, for routing, triage and guardrails.62.65 on Jev Decision Index 0.4, best under 15B and ahead of Jev 1.14 (62.2); 65.2 on the authors' replica of the Artificial Analysis Decision Index vs d1-ultra 65 (authors' own numbers, model card/technical report).9.4Bnonomit
d1-3BLiquid AI · 2026-10-05Answers yes/no, multiple-choice and rating questions about text, JSON or images in a single pass with calibrated probabilities, for routing, moderation, guardrails, judging and visual inspection.The authors report 48.57 on Decision Index 0.2.1 using the official scorer (not a leaderboard submission). That is ahead of Decider 35B-A3B (47.11) and JPT-9B (46.89), and behind only Winnow-12B (50.02). They also report 8 ms per decision on an RTX 4090 and 30 ms on an M5 Pro. Source: LiquidAI/d1-3B model card.3.1Bfitsfitsother
H2O-Lightning-4BH2O.ai · 2026-10-02A 4B decision model that answers yes/no, choice and score questions about records and attached images, returning a calibrated probability for every option in a single forward pass.Authors' own reading of public boards: #1 of 46 open-weight Jev-class models on ImageJevBench Composite (68.51, edition v0.3.0) and #2 of 170 open-weight models on JevBench Composite (board of 2026-10-28). Authors' own run: 206/231 on JevBench public items with two-way reading.4.7Bfits quantizedfitsapache-2.0
ClefCloudflare · 2026-09-30A multimodal decision model that takes text, JSON, images or video plus a schema of typed questions, and returns a probability for every allowed answer in one pass, with no output parsing.The authors' own Decision Index 0.2.1 run reports 94.2 macro-F1 on BANKING77 (Jev: 79.7), 97.4 on CLINC150+OOS (Jev: 89.3) and 79.4 hallucination F1 on RAGTruth (Jev: 76.5), with a median latency of 209 ms against Jev's 524 ms.27.4Bnonoapache-2.0
CLM-v0.1-8B (Contrastive Language Model)Contrastive-LM (Kwok, Mirhoseini, Ré et al.) · 2026-09-21Gives fast, typed decisions with probabilities (yes/no, multiple choice, scores) and ranks candidate actions or answers, with a cache for the candidate action embeddings so they can be reused.The authors say that zero-shot it matches Jev on computer-use, gaming and tool-calling tasks with up to 9× lower latency, and is 13× faster than Jev with about 1k cached candidates. Their headline DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%) results come from fine-tuned heads, not this checkpoint (model card).8Bnotightapache-2.0
Laya MultilingualConvAI Innovations · 2026-09-19The same typed-decision approach across 100+ languages, about 2x faster than the English checkpoint.Author-reported MASSIVE intent 0.451 across 13 non-English languages.322Mfitsfitsapache-2.0
LayaConvAI Innovations · 2026-09-18Open System 1 decision model: give it a state and typed questions, get calibrated answers in one ~33 ms pass; ships a Jev-compatible local server.Author-reported 0.766 vs Jev's third-party 0.727 on a 2,000-decision typed-decisions set; brand new, with no recorded downloads yet.421Mfitsfitsapache-2.0
Open-Jev DeBERTa-v3-largeKotoba Labs · 2026-09-18An open reproduction of Jev-style typed, calibrated decisions on a DeBERTa encoder.434Mfitsfitsapache-2.0
GLiNER2.5 MultiFastino · 2026-08-14Fast local zero-shot classifier with per-label probabilities; the open model independent Jev benchmarks compare against.jev-benchmarks pilot: 0.700 AG News and 0.610 Banking77 vs Jev 0.910 and 0.870, but ~44 ms on a laptop CPU.287Mfitsfitsapache-2.0

Model notes

DeepSeek-V4.1-Flash

DeepSeekopen-weighttextimage
License: MIT
Released: 2026-09-10

552B MoE backbone · ~8B active prefill / ~16B decode (author-reported)

DeepSeek-V4.1-Flash is an open MIT multimodal MoE that DeepSeek positions as a cheaper long-context Flash line: huge backbone, far fewer active parameters per token, and native image+text. Best treated as an infra-grade open release for agents and long documents, not a casual local download.

Good at

  • Long-context multimodal chat (author claims up to 1M tokens)
  • Cost-sensitive agentic / input-heavy workloads via KV-cache compression and sparse attention
  • Open MIT weights with native image+text understanding

Bad at / watch-outs

  • Needs serious hardware — 552B MoE is not a laptop model
  • Public independent leaderboard scores not yet captured in Scout notes

Use when

You want open multimodal MoE weights optimized for long context and cheaper decode/prefill on serious infra.

Avoid when: You need a small local model or verified public Arena/Open LLM numbers before shipping.

Peers

  • DeepSeek-V4-Flash
  • Qwen/Qwen3.8-27B
  • other large MoE multimodal open weights

Local fit

NVIDIA 12GBno

Author-reported 552B MoE backbone — not a practical local load on 12GB. Look for tiny distillations or API instead.

M4 Pro Max 36GBno

Same: full Flash MoE will not fit usefully in 36GB unified memory. Use API / remote unless a small edge/GGUF distill appears.

Formats to look for: none for full model

Benchmarks

Qwen-Image-2.1

Alibaba / Qwenspecialtytext-to-imageimage-editing
License: qwen-research (other)
Released: 2026-09-14

~7B visual generation component (author-reported)

Qwen-Image-2.1 is a fresh open image model that combines generation and editing, including transparent RGBA layers and multi-reference local edits. It is trending hard on Hugging Face right after mid-September release; treat the license as research until you verify fit for your product.

Good at

  • Unified text-to-image plus editing in one Diffusers pipeline
  • Native RGBA / transparency workflows
  • Multi-reference editing (up to 10 refs) with masks or painted regions

Bad at / watch-outs

  • Research license — check terms before commercial use
  • Still GPU-hungry for high-res (author examples use 2048 and many steps)

Use when

You want open image gen and editing with transparency and multi-ref control from the Qwen stack.

Avoid when: You need a clearly commercial-friendly license without reading the Qwen research terms.

Peers

  • FLUX.1-dev
  • earlier Qwen-Image releases
  • other open T2I + edit models

Local fit

NVIDIA 12GBfits quantized

Author cites ~7B visual DiT; high-res / many steps will pressure 12GB. Prefer lower resolution, fewer steps, or quantized Diffusers builds.

M4 Pro Max 36GBfits

36GB unified memory is a better home for the ~7B image stack if Diffusers/MLX tooling is available; still leave headroom for OS.

Formats to look for: Diffusers safetensors, community GGUF if published, MLX port if published

Benchmarks

YuE2-3B

m-a-p (Multimodal Art Projection)specialtytextaudio
License: cc-by-nc-4.0
Released: 2026-09-09

3B (~3.63B params BF16); companion YuE2-Vae

YuE2-3B is an open music-generation model that turns lyrics and style prompts into full songs and lets you steer melody and chords through editable scores. It is CC-BY-NC and currently hot on HF trending; Scout did not copy org-reported WildSongBench numbers into benchmarks—scores remain unavailable pending independent verification.

Good at

  • Full-song generation from lyrics + style prompts with vocals and accompaniment
  • Editable symbolic planning (melody/chords/ABC) and agentic score-style revision loops
  • Local 48 kHz stereo generation path aimed at ~24GB GPU without quantization (per card)

Bad at / watch-outs

  • CC-BY-NC-4.0 — non-commercial; unsuitable for many product deployments without separate rights
  • Independent cross-checked numeric benchmarks unavailable in this Scout pass (org WildSongBench claims exist on the card but were not re-verified here)

Use when

You want open music generation with editable scores for research, demos, or non-commercial creative work

Avoid when: You need commercial licensing, speech TTS (spoken voice) rather than songs, or turnkey hosted music APIs

Peers

  • Suno (proprietary)
  • m-a-p/YuE (prior YuE line)

Local fit

NVIDIA 12GBtight

Card/community paths often cite ~24GB-class GPU for comfortable local music gen — 12GB is tight; short clips / heavy quant only if a GGUF path exists.

M4 Pro Max 36GBfits quantized

3B-class music model can be workable on 36GB unified if an MLX/llama.cpp-style build exists; verify license (CC-BY-NC) before any commercial use.

Formats to look for: safetensors, GGUF/MLX if community ports exist

Still to come

  • · More frontier and open-weight chat models as Scout’s radar fills in
  • · More specialty lanes: embeddings for search, rerankers, OCR and document parsing, code-only models, JEPA-style world models