AI Models

Anthropic's mid-tier model: close to Opus 5.5 on many benchmarks at half the price, $2/$10 per million tokens.

GPT-6.1 Sol

OpenAIGeneral purpose

OpenAI's mid-priced workhorse: near-Astra results on coding and computer use at $2/$10 per million tokens.

Nano Banana 2.1

GoogleImage generation

Google's Flash-tier image model: fourth in blind votes, from about 3.4 cents per 1K image, with up to 14 reference images.

Voyage rerank-3

Voyage AI (MongoDB)Embeddings & reranking

Voyage AI's newest rerankers, with a 32K-token window, instruction following and prices from $0.02 per million tokens.

Claude Haiku 5.5

AnthropicGeneral purpose

Anthropic's small, fast model, from $0.10/$0.50 per million tokens, with a 1M-token context and adjustable effort.

Gemini 3.8 Flash TTS

GoogleText to speech

Google's expressive text-to-speech, fifth in blind votes at about $16.50 per million characters, with voice design and cloning.

Gemini 4 Argon

GoogleReasoning

Google's new flagship reasoning model: reads text, images, audio, video and PDFs, with a 1M-token context.

MiMo-V2.6-Pro

XiaomiReasoningOpen weights

Xiaomi's 1-trillion-parameter open-weight model, top of the open models on Artificial Analysis, at $0.87 per 1M output.

NVIDIA Nemotron 3.5 Transcribe

NVIDIASpeech to textOpen weights

NVIDIA's promptable open speech model: transcription in 40 languages, speaker labels and translation in one checkpoint.

Perplexity Decider (pplx-decider)

PerplexityDecision modelsOpen weights

Perplexity's open 27B and 4B decision models, also sold as a $0.02-per-million-token API: text and images in, probabilities out.

Qwen3.8-27B

Alibaba (Qwen)General purposeOpen weights

Alibaba's Apache 2.0 dense model: 27 billion parameters, image and video input, and coding scores that rival models many times its size.

voyage-4-large

Voyage AI (MongoDB)Embeddings & reranking

Voyage AI's top text embedding model, built on mixture-of-experts, sharing one vector space with cheaper Voyage 4 models.

Wan 3.0

AlibabaVideo

Alibaba's newest video model, top of the current Artificial Analysis video board, with clips up to 30 seconds.

Cartesia Sonic 3.6

CartesiaText to speech

Cartesia's low-latency text-to-speech model for voice agents, in 44 languages.

Cohere Rerank 5

CohereEmbeddings & reranking

Cohere's October 2026 reranker in Pro and Fast sizes, with calibrated scores and an option to return only the relevant spans.

ElevenLabs Scribe v2

ElevenLabsSpeech to text

ElevenLabs' batch transcription model: 2.2% word error rate on independent tests, $0.22 per hour, with diarization and entity detection.

Gemma 4

GoogleSmall & on-deviceOpen weights

Google's Apache 2.0 open model family, from phone-sized E2B to a 31B model for one GPU, with image, video and audio input.

GLM-5.3

Z.ai (Zhipu AI)CodingOpen weights

Z.ai's open-weight coding and agent model, near-MIT licensed, at $1.40/$4.40 per million tokens on its API.

PaddleOCR-VL-1.6

Baidu (PaddlePaddle)OCR & documentsOpen weights

A 0.9-billion-parameter document parser that tops the main OCR benchmark, at least on its makers' count.

SAM 3 (Segment Anything Model 3)

MetaVision & multimodalOpen weights

Meta's 848M-parameter model that finds, outlines and tracks every instance of a short text phrase in images and video.

Seedance 2.5

ByteDanceVideo

ByteDance's joint audio-video model with 30-second takes, heavy multimodal referencing and timestamp edits.

Voxtral 3 (Transcribe and TTS)

Mistral AISpeech to text

Mistral's October 2026 speech family: transcription in 50 languages from $0.0035 a minute, real-time under 500 ms, and voice cloning.

Claude Fable 5.1

AnthropicReasoning

Anthropic's top tier for hard reasoning and long-horizon agent work, priced at $10/$50 per million tokens.

DeepSeek-V4-Pro

DeepSeekReasoningOpen weights

DeepSeek's largest open-weight model: 1.6T parameters, MIT licence, 1M-token context and very low API prices.

EmbeddingGemma 2

Google DeepMindEmbeddings & rerankingOpen weights

Google's open, on-device embedding model that puts text, code, images, audio and video in one space.

Gemini 4 Carbon

GoogleReasoning

Google's top reasoning model, first on Artificial Analysis's index, but for now only vetted cyber defenders can use it.

Gemini Embedding 2

GoogleEmbeddings & reranking

Google's first multimodal embedding model: text, images, video, audio and PDFs in one shared vector space.

Gemini Omni Flash 1.1

GoogleVideo

Google's default video model, built on Gemini, for short clips you refine through conversation.

GLM-5.3-Flash

Z.ai (Zhipu AI)General purposeOpen weights

Z.ai's MIT-licensed 320B mixture-of-experts model with image and video input, a 1M-token context and a $0.15/$0.50 API price.

GLM-OCR

Z.ai (Zhipu AI)OCR & documentsOpen weights

Zhipu's compact OCR model: independently scored, 1.74 million downloads a month, and an API at $0.03 per million tokens.

GPT-6 Luna

OpenAIGeneral purpose

OpenAI's small, cheap model for high-volume tasks, at $0.10/$0.50 per million tokens with a 1.05M-token context.

GPT-Live-1 (and GPT-Realtime-2.1)

OpenAIMusic & audio

OpenAI's full-duplex voice model: it listens while it talks and hands the thinking to a text model you choose, at $0.05 a minute.

Jina Embeddings v5

Jina AI (Elastic)Embeddings & rerankingOpen weights

Jina AI's small multilingual text embedders and their omni siblings for images, video and audio, sharing one vector space.

Jina Reranker v3.5

Jina AI (Elastic)Embeddings & rerankingOpen weights

Small 0.6B multilingual reranker from Jina AI (now Elastic) that scores a whole candidate list in one pass, with 131K-token context.

Kimi K3

Moonshot AIReasoningOpen weights

Moonshot AI's 2.8-trillion-parameter open-weight model, strong on agent and browsing tasks, with a 1M-token context.

MAI-Image-2.6

Microsoft AIImage generation

Microsoft's in-house image model, fifth on Artificial Analysis's arena, in preview on Foundry at about $39 per 1,000 images.

Midjourney V8

MidjourneyImage generation

Midjourney's eighth-generation image model, now at V8.2 with an instruction editor, sold only as a monthly subscription.

MiniCPM-OCR 1.0

OpenBMBOCR & documentsOpen weights

OpenBMB's 1.3B OCR model, fastest in its maker's 15-model test and top on four of six OCR benchmarks, under Apache 2.0.

MiniMax H3 (Hailuo 3.0)

MiniMaxVideoOpen weights

MiniMax's 33B video model with stereo sound: the top open-weights entry, licensed out of the US, EU, UK and South Korea.

Qwen3-VL-Embedding

Alibaba (Qwen)Embeddings & rerankingOpen weights

Alibaba's open multimodal embedder for text, images, screenshots and video in one vector space, in 2B and 8B sizes.

Chandra OCR 2

DatalabOCR & documentsOpen weights

Datalab's Qwen 3.5-based parser leads olmOCR-Bench and an independent historical-print board, but its weights are free only under $2M.

Cohere Embed 5

CohereEmbeddings & reranking

Cohere's multimodal embedding models for text, images and PDFs, with a 128K-token input and Pro and Fast tiers that share one vector space.

DeepSeek-V4.1-Flash

DeepSeekGeneral purposeOpen weights

DeepSeek's open MIT-licensed mixture-of-experts model with image input and a 1M-token context, from $0.15 per million input tokens off-peak.

Fish Audio Drama 3

Fish Audio (39 AI, Inc.)Text to speechOpen weights

Fish Audio's 35B open-weight speech model for directed, multi-character scenes: the top open model on Artificial Analysis's TTS board.

FLUX 3

Black Forest LabsVideo

Black Forest Labs' multimodal model for video with audio and images, built on one shared architecture.

Gemini 3.5 Transcribe

GoogleSpeech to text

Google's dedicated speech-to-text models, with a streaming Live variant, at about half a cent per minute.

GPT-6 Astra

OpenAIReasoning

OpenAI's flagship reasoning model for the hardest work, at $10/$50 per million tokens with a 1.05M-token context.

Hy4 preview

Tencent (Hy, formerly Hunyuan)CodingOpen weights

Tencent's 770B open-weight flagship (formerly Hunyuan) under plain Apache 2.0, strong at coding with a 1M-token context.

Kimi K3.1 Pro

Moonshot AIReasoning

Moonshot AI's post-trained Kimi K3: top of Artificial Analysis's Terminal-Bench table, but 22nd on its harder v5.0 index, and slow.

LightOnOCR-3-4B

LightOnOCR & documentsOpen weights

LightOn's 4B end-to-end OCR model with layout boxes and chart data, first on ParseBench and best on French documents by its maker's tests.

Mistral OCR 4.1

Mistral AIOCR & documents

Mistral's document OCR with paragraph bounding boxes, block labels and confidence scores in 170 languages, at $4 per 1,000 pages.

Nano Banana Pro (Gemini 3 Pro Image)

GoogleImage generation

Google's premium image model: four times the price of Nano Banana 2.1, but still the one reviewers trust for realism.

OvisOCR3

Alibaba (ATH-MaaS)OCR & documentsOpen weights

OvisOCR2's successor: the same 0.8B size, now with field extraction and cross-page tables, on scores only Alibaba has checked so far.

Qwen3-Embedding and Qwen3-Reranker

Alibaba (Qwen)Embeddings & rerankingOpen weights

Alibaba's Apache 2.0 embedding and reranking family, from a 0.6B model to an 8B that topped MTEB multilingual in June 2025.

Qwen3.8-Flash-Next

Alibaba (Qwen)General purposeOpen weights

Qwen's preview of the Qwen4 architecture: 6 billion active parameters, strong vendor-reported coding scores and a restrictive licence.

Qwen3.8-Max

Alibaba (Qwen)ReasoningOpen weights

Alibaba's 2.4-trillion-parameter flagship, with open weights and a $2/$6 API that adds image, video and 1M context.

Shieldstral 1.0 3B

Mistral AIDecision modelsOpen weights

Mistral's open 3B safety classifier that checks text and images against moderation policies you write in plain language.

Step 5 Preview

StepFunReasoningOpen weights

StepFun's 600B open-weight model with a 1M-token context, as smart as Kimi K3 on Artificial Analysis for half the cost.

Unlimited-OCR 2

BaiduOCR & documentsOpen weights

Baidu's open OCR model that parses a 300-page document in one request, with memory use that stays flat as the document grows.

d3-turbo (vLLM Semantic Router)

vLLM Semantic RouterDecision modelsOpen weights

vLLM Semantic Router's upgraded 27B decision model: reads an input once, answers up to 255 questions, and locates objects in video.

Doubao Seed 2.1 Pro

ByteDanceGeneral purpose

ByteDance's closed flagship behind the Doubao app: a leader on video and visual understanding, sold mainly through China's Volcano Engine.

Gemini 3.5 Flash-Lite

GoogleGeneral purpose

Google's cheapest current Gemini text model, very fast once it starts, at $0.30/$2.50 per million tokens and with a free tier.

Gemini 3.8 Flash

GoogleGeneral purpose

Google's mid-tier model for agents and documents, reading text, images, audio, video and PDFs with a 1M-token input.

HunyuanOCR 1.5

Tencent (Hy, formerly Hunyuan)OCR & documentsOpen weights

Tencent's 1.12B end-to-end OCR model for parsing, spotting, extraction and translation; not licensed in the EU, UK or South Korea.

HY-World 2.0

Tencent (Hy, formerly Hunyuan)Game & world modelsOpen weights

Tencent's open pipeline that turns a sentence or photo into a persistent 3D world you can walk through and open in Unity.

Kling 4.0

Kuaishou (Kling AI)Video

Kuaishou's new video model: 30-second takes, 10 keyframes, 15 references and stereo sound, with an API so far only on fal.

Ministral 3 (3B, 8B and 14B)

Mistral AISmall & on-deviceOpen weights

Mistral's small open models in three sizes, each with base, instruct and reasoning versions and image input, under Apache 2.0.

Moondream Parakeet Redux

Moondream (M87 Labs)Speech to textOpen weights

Moondream's 1.58-bit Parakeet: NVIDIA's 25-language transcriber shrunk from 1.2 GB to 178 MB and run at 113x real time on eight CPU cores.

Muse Glimmer

MetaGeneral purposeOpen weights

Meta's 30B Apache 2.0 agent model, distilled from Muse Spark and packaged with tested 4-bit builds for one consumer GPU.

Muse Wonder

MetaGeneral purpose

Meta's first frontier-priced model: strong on long coding and legal agent work, behind Claude Opus 5.5 on most of Meta's own table.

Nemotron 3.5 Super VL

NVIDIAVision & multimodalOpen weights

NVIDIA's open 120B hybrid Mamba MoE that reads images, documents and video, with a 1M-token context and published training recipes.

OvisOCR2

Alibaba (ATH-MaaS)OCR & documentsOpen weights

Alibaba's 0.8B page parser, near the top of the independent OmniDocBench table, already replaced by OvisOCR3.

Parakeet TDT 0.6B v3

NVIDIASpeech to textOpen weights

NVIDIA's 600M-parameter transcription model for 25 European languages: the fastest on Artificial Analysis's board.

Phonon-2

Fermion ResearchSpeech to textOpen weights

NVIDIA's Parakeet speech model squeezed into a 164MB download with almost all of its English accuracy, for laptops and phones.

PP-OCRv6

Baidu (PaddlePaddle)OCR & documentsOpen weights

Baidu's classic detect-then-read OCR pipeline in three sizes from 1.5M to 34.5M parameters, under Apache 2.0.

pyannote speaker-diarization-community-1

pyannoteAISpeech to textOpen weights

The open speaker-diarisation pipeline most transcription stacks use to work out who spoke when, now with fewer speaker mix-ups.

Seedream 5.0 Pro

ByteDanceImage generation

ByteDance's flagship image model, built for editable design work: up to 20 layers, coordinate edits and text in 14 languages.

Xiaomi-OCR-0

XiaomiOCR & documentsOpen weights

Xiaomi's 0.8B open OCR model that parses pages, answers questions about them and fills JSON fields, under Apache 2.0.

d3 (vLLM Semantic Router)

vLLM Semantic RouterDecision modelsOpen weights

vLLM Semantic Router's open 27B decision model with video input and stock vLLM serving, already overtaken by its own d3-turbo.

Deepgram Nova-3

DeepgramSpeech to text

Deepgram's speech-to-text model: second-fastest on Artificial Analysis's test and cheap per minute, but well off the accuracy leaders.

Falcon OCR

Technology Innovation Institute (TII)OCR & documentsOpen weights

TII's 270M-parameter OCR model reads a page in one small Transformer and gets surprisingly close to models three to ten times its size.

FLUX 3 Image Klein

Black Forest LabsImage generationOpen weights

Black Forest Labs' 5.6B open image model with FLUX 3's 4K output, editing and bounding boxes, free to use commercially under $5M revenue.

GPT-5.6-Cyber

OpenAIReasoning

OpenAI's restricted cybersecurity model for vetted researchers: it writes exploits and tests defences that its standard models refuse.

gpt-oss-120b

OpenAIReasoningOpen weights

OpenAI's larger open-weight reasoning model: Apache 2.0, 5.1B active parameters, and small enough for a single 80GB GPU.

GPT-Rosalind

OpenAIReasoning

OpenAI's restricted life-sciences model for approved research organisations: drug discovery, genomics and protein work.

Granite 4.2

IBMReasoningOpen weights

IBM's Apache 2.0 family of 3B, 8B and 30B dense reasoning models for agents and coding, released in August 2026.

Ideogram 4.0

IdeogramImage generationOpen weights

Ideogram's 9.3B design model, strong on in-image text; its downloadable weights are non-commercial, so paid work goes through the API.

LFM3

Liquid AISmall & on-deviceOpen weights

Liquid AI's family of six 2B models for text, vision, audio, omni and typed decisions, all grown from one 1.7B base.

Ling-3.1-flash

Ant Group (inclusionAI)ReasoningOpen weights

Ant Group's 560B open-weight model: fast at over 200 tokens a second, MIT-licensed, with a 1M-token context.

Maya 2 Native

Maya ResearchText to speechOpen weights

A 2.5B Apache 2.0 streaming voice for 11 Indian languages and English, level with Mistral and Eleven v3 on VoiceArena's Hindi board.

MiniCPM-V 4.7 (35B-A3B and 1B)

OpenBMBVision & multimodalOpen weights

OpenBMB's Apache 2.0 vision models: a 35B mixture-of-experts that leads on documents and a 1B for phones.

MiniCPM5-2B

OpenBMBSmall & on-deviceOpen weights

OpenBMB's 2.5B on-device model: tool calls and code fixes like a 4B model, with its checkpoints and training data published.

MiniMax-M3.1-Flash

MiniMaxGeneral purposeOpen weights

MiniMax's always-reasoning open-weight MoE: M3's 428B design, a flat $0.15/$0.60 API, 1M context and a hallucination problem.

Muse Spark 1.3

MetaGeneral purpose

Meta's closed agent and coding model: a 1M-token context at $1.25/$4.25, now the budget tier beneath Muse Wonder.

Muse Spark Open

MetaGeneral purposeOpen weights

Meta's 816B open-weight Muse Spark: a 1M-token agent model you can self-host, under a licence with strings attached.

OmniVoice

k2-fsa (Next-gen Kaldi)Text to speechOpen weights

Open zero-shot TTS with voice cloning and voice design in 600+ languages, from the team behind Next-gen Kaldi.

Qianfan-OCR

BaiduOCR & documentsOpen weights

Baidu's 4.7B model that parses pages, extracts fields and reads charts in one pass, at a cost in size and page accuracy.

Qwen-Image-2.1

Alibaba (Qwen)Image generationOpen weights

Alibaba's 7B open image generator and editor tops the open-weights board and outputs transparency, but is licensed for research only.

Surya OCR 2.5

DatalabOCR & documentsOpen weights

Datalab's 689M-parameter document parser: layout, OCR, tables and maths as HTML, fast on one GPU, free under $5M revenue.

Ternary Bonsai 2 9B

PrismMLSmall & on-deviceOpen weights

PrismML's ternary rebuild of Qwen3.5-9B: a vision-language model with tool calling in a 2.42 GB file that runs on phones and laptops.

Codestral 25.08

Mistral AICoding

Mistral's fast fill-in-the-middle model for code autocomplete, still a default in IDE tools; not built for agent work.

DeepSeek-V4.1-Flash-1017 GSQ-RCO Coder (GGUF)

ISTA-DASLab (Institute of Science and Technology Austria)CodingOpen weights

DeepSeek-V4.1-Flash for code, with half its experts cut and the rest at about 2 bits: a 72.1 GB transformer.

FireRed-OCR

Xiaohongshu (FireRedTeam)OCR & documentsOpen weights

Xiaohongshu's 2B fine-tune of Qwen3-VL leads the end-to-end OCR models on its own table, and trails the pipelines there too.

Flint

RecursiveCodingOpen weights

Recursive's first model: an open-weight coding model with 40B parameters, 2.5B active, built to run fast on a laptop.

GPT Transcribe

OpenAISpeech to text

OpenAI's replacement for Whisper: more accurate and cheaper than GPT-4o Transcribe, but without timestamps or speaker labels.

Kokoro-82M

hexgradText to speechOpen weights

An 82M-parameter Apache 2.0 speech model, mid-table on quality and the cheapest voice on Artificial Analysis's board.

Kolibri 1

Aleph AlphaReasoningOpen weights

Aleph Alpha's Apache 2.0 German-English reasoning model: 78B parameters, 3.46B active, trained in Germany and Finland.

Luma Ray3.2

Luma AIVideo

Luma's video model for post-production: 16 keyframes, native HDR and 16-bit EXR, sold per clip through the API and the Luma app.

Mage-VL-Cowork

MicrosoftVision & multimodalOpen weights

Microsoft's 4B desktop agent that chooses, step by step, between clicking the screen and running code in a terminal.

Marengo 3.5

TwelveLabsEmbeddings & reranking

TwelveLabs' video-first embedding model: video, audio, images, text and PDF pages in one 512-dimension space, billed per token.

Matrix-Game 4.0

Skywork AI (Kunlun Tech)Game & world modelsOpen weights

Skywork's open 5B model that repaints a game engine's low-poly frames in real time, for its three bundled scenes only.

MiniMax-M3

MiniMaxGeneral purposeOpen weights

MiniMax's open-weight multimodal MoE: 1M-token context, sparse attention, a $0.30/$1.20 API and a licence with strings attached.

Mistral Large 4

Mistral AIGeneral purpose

Mistral's 1-trillion-parameter flagship, trained and hosted in Europe, in preview at $1.36/$4.18 with weights promised.

MOSS-Transcribe-Diarize 0.9B

OpenMOSS (MOSI)Speech to textOpen weights

A 0.9B open model that transcribes, labels speakers and timestamps up to 90 minutes of audio in one pass.

Nemotron 3.5 Bolt (74B-A9B 24GB and E4B 4GB)

NVIDIASmall & on-deviceOpen weights

NVIDIA's local-agent pair: a 2-bit 74B vision model named for 24 GB and a Gemma-based 4B for voice assistants.

Ripple-65B Preview

Ricursive IntelligenceCodingOpen weights

A 65B dense open-weight coding model from the AlphaChip founders, with self-reported agentic scores level with far larger rivals.

Sage-A1.5

Shanghai AI Laboratory (InternScience)ReasoningOpen weights

Shanghai AI Lab's 35B agent model for research tasks, with built-in sub-agent fan-out and a self-check mode, under Apache 2.0.

Sakana Fugu (Fugu, Max and Ultra v2)

Sakana AIReasoning

Sakana AI's orchestrator: one API that splits each task across a hidden pool of frontier models and merges the answers.

zerank-2

ZeroEntropy (Notion)Embeddings & rerankingOpen weights

ZeroEntropy's 4B open reranker with calibrated scores, Apache 2.0 since Notion bought the company and its hosted API went dark.

Audio8 TTS

Edge0Text to speechOpen weights

A 0.6B-parameter voice-cloning model in 23 languages under Apache 2.0, with error rates close to models several times its size.

Beam

ReflectionReasoningOpen weights

Reflection's first open model: 501B parameters, 23B active, Apache 2.0, and fewer reasoning tokens per task than its rivals.

Breeze TTS 2

BreezeBlueText to speechOpen weights

A 3.5B-parameter voice cloning and voice design model that sits near the top of the open-weights arena, but bars commercial self-hosting.

Command A+

CohereGeneral purposeOpen weights

Cohere's 218B mixture-of-experts model for agents and multilingual RAG, with Apache 2.0 weights that fit on two H100s.

gpt-oss-20b

OpenAISmall & on-deviceOpen weights

OpenAI's Apache 2.0 open-weight reasoning model with 3.6B active parameters, sized to run in 16GB of memory on a laptop or consumer GPU.

Grok 4.7

SpaceXSI (formerly xAI)General purpose

SpaceXSI's flagship Grok model for chat, code and agents: $2/$6 per million tokens, a 500K context and very long thinking times.

Infinity-Parser2-Flash

INF Technology (infly)OCR & documentsOpen weights

INF's 2.2B parser, strong on olmOCR-Bench and charts, but slow on a single stream and weaker on OmniDocBench than smaller rivals.

Jev

TypeSafe AIDecision models

TypeSafe AI's hosted decision model: typed choices, scores and yes/no probabilities for $0.042 per million input tokens, output free.

LFM2.5-24B-A2B

Liquid AISmall & on-deviceOpen weights

Liquid AI's 24B reasoning model with 2.45B active parameters, trained inside agent harnesses to call tools on a laptop.

Ling-3.0-flash-Sante

Ant Group (inclusionAI)General purposeOpen weights

Ant Group's MIT-licensed 124B medical model, 5.1B active: near frontier on its own health benchmarks, at a fraction of the price.

LTX-2.5

Lightricks (LTX)VideoOpen weights

Lightricks' 22B open video model with synchronised sound and multishot scenes, up to 4K on its API from $0.09 a second.

Mistral Medium 3.5

Mistral AIGeneral purposeOpen weights

Mistral's 128B dense open-weight coding and agent model fits four GPUs, but its licence shuts out firms above $20M a month.

MOSS-Image

OpenMOSS (MOSI)Image generationOpen weights

A 5B open image model under Apache 2.0 that places text and objects where a JSON layout tells it to, and renders long text well.

mxbai-rerank v2

MixedbreadEmbeddings & rerankingOpen weights

Mixedbread's Apache 2.0 rerankers in 0.5B and 1.5B sizes, with 100+ languages and an 8K-token context, built to self-host.

Nanbeige4.3-33B

BOSS Zhipin (Nanbeige LLM Lab)CodingOpen weights

A dense 33B open model from BOSS Zhipin's lab that claims to lead the 30B class at agentic coding, under Apache 2.0.

Nemotron ColEmbed V2

NVIDIAEmbeddings & rerankingOpen weights

NVIDIA's late-interaction retrievers that search page images without OCR: accurate, storage-hungry and non-commercial at every size.

Nemotron OCR v2

NVIDIAOCR & documentsOpen weights

NVIDIA's 84M-parameter detector-recogniser OCR for RAG ingestion: about 35 pages a second on an A100, licensed for commercial use.

Pika H3 Flex

Pika (Mellis, Inc.)VideoOpen weights

Pika's six-step post-train of MiniMax H3: video with stereo sound at half MiniMax's API price, under MiniMax's territorial licence.

Runway Gen-4.5

RunwayVideo

Runway's in-house video model: a 2025 leaderboard winner, now outranked, sold inside a suite that also resells its rivals.

VibeVoice-ASR

MicrosoftSpeech to textOpen weights

Microsoft's MIT-licensed transcriber that takes an hour of audio in one pass and labels who spoke when.

Whistle

Cactus ComputeSpeech to textOpen weights

A 16.9 MB speech-to-text model in seven languages that runs on a CPU with no dependencies; a community port fits it on a $10 chip.

ZDTaichu5.0-9B

Zidong Taichu (CASIA and Wuhan AI Research)Vision & multimodalOpen weights

A 9B vision-language model from China's Zidong Taichu team that trades a little OCR for spatial reasoning and tool use.

Chatterbox

Resemble AIText to speechOpen weights

Resemble AI's MIT-licensed voice-cloning models, from a 110M CPU model to 23 languages, with a watermark on every clip.

Cohere Rerank 4

CohereEmbeddings & reranking

Cohere's December 2025 reranker in Pro and Fast versions with a 32K-token context, replaced by Rerank 5 on 21 October 2026.

DecisionNode

Bynn IntelligenceDecision models

Bynn Intelligence's hosted decision model, returning six typed answers with confidence, plus two small open-weight Edge models for devices.

DeepSeek-OCR 2

DeepSeekOCR & documentsOpen weights

DeepSeek's 3B open document reader reorders a page before reading it, but PaddleOCR-VL and Unlimited-OCR now outscore it.

ERNIE 5.1

BaiduGeneral purpose

Baidu's closed flagship, distilled from ERNIE 5.0 at a fraction of the cost: strong at search and maths, hard to buy outside China.

FrogNano-4B-2609

MicrosoftCodingOpen weights

Microsoft's 4B research coding agent, trained only by reinforcement learning on synthetic tasks: 61.5% on SWE-bench Verified, self-reported.

humanizer (12B)

jialinyyzz (independent developer)Small & on-deviceOpen weights

An open 12B Gemma 4 fine-tune that rewrites AI-written English and Chinese drafts to read as human, offline, keeping the facts.

Hy-Memory-30B-A3B

Tencent (Hy, formerly Hunyuan)General purposeOpen weights

Tencent's 3B-active model that does one job: deciding what an agent's long-term memory should keep, in place of a frontier LLM.

Iris-3B

SperidlabsImage generationOpen weights

Speridlabs' 3B text-to-image model that paints pixels directly, with no VAE, plus a 4-step Turbo and depth and upscaling fine-tunes.

Jina OCR v1

Jina AI (Elastic)OCR & documentsOpen weights

A DeepSeek-OCR fine-tune from Jina AI, now part of Elastic, tuned for pages per second on cheap GPUs; weights are non-commercial.

Mistral Small 4

Mistral AIGeneral purposeOpen weights

Mistral's Apache 2.0 mixture-of-experts: 119B parameters, image input, switchable reasoning, $0.15/$0.60 per million tokens.

ScaleDown Constrained Decoding Models (Classify, Extract, Summarize)

ScaleDownDecision models

ScaleDown's hosted classify, extract and summarise models at $0.05 per million input tokens, with a Jev-shaped Decisions endpoint.

Spark-X2.6-4B

iFlytek (SparkLLM)Small & on-deviceOpen weights

iFlytek's 4B on-device agent model under Apache 2.0, claiming to beat Qwen3.5-9B on tools and code, with a 1M-token context.

text-embedding-3-large

OpenAIEmbeddings & reranking

OpenAI's top embedding model, unchanged since January 2024: supported everywhere and cheap, but no longer a retrieval leader.

Whisper large-v3-turbo

OpenAISpeech to textOpen weights

OpenAI's open-weights speech recognition model, still the default for self-hosted transcription.

Z-Image-Turbo

Alibaba (Tongyi-MAI)Image generationOpen weights

A 6B Apache-licensed image model from Alibaba's Tongyi lab that runs in eight steps on a 16GB consumer GPU.

Aion 3.5

Aion Labs (Deep Forge)Role-play & characters

A role-play and storytelling system in which several GLM-based models write each reply together; no public benchmarks yet.

Clef

CloudflareDecision modelsOpen weights

Cloudflare's open 27B decision model that scores typed choices about text and images, served on Workers AI behind a Jev-compatible API.

Clover-35B-A3B-Preview

DiffuSpaceGeneral purposeOpen weights

A diffusion language model converted from Qwen3.6-35B-A3B that writes 32-token blocks in parallel, open under Apache 2.0.

dots.ocr

Xiaohongshu (rednote hi lab)OCR & documentsOpen weights

Xiaohongshu's 2025 multilingual layout parser is still heavily downloaded, but its own makers and three newer rivals have moved past it.

Falcon-H1R-7B

Technology Innovation Institute (TII)ReasoningOpen weights

TII's 7B maths and coding reasoning model with a hybrid Transformer-Mamba design, a 256K context and an attribution licence.

GLiNER2.5-Decide

FastinoDecision modelsOpen weights

Fastino's sub-billion-parameter zero-shot classifier: pass any labels at call time and get one back on a CPU, no prompt or generation.

Limite 1B - Violetto

ParadigmaSmall & on-deviceOpen weights

A 1B maths model from a Rome start-up that matches 30B models on competition problems and misreads almost anything else.

Needle 3

Cactus ComputeSmall & on-deviceOpen weights

A 29M to 121M model that turns requests into function calls on phones, robots and hubs; strong on phone commands, weaker elsewhere.

Ornith-1.5-122B

OrnithCodingOpen weights

A 122B open-weight reasoning and coding model trained by a self-improvement loop that writes its own tasks, under MIT.

Qwen-Image-Edit-2511

Alibaba (Qwen)Image editingOpen weights

Alibaba's Apache-licensed image editor: mid-table on quality, but the cleanest licence among open editors you can ship.

Qwen3.5-397B-A17B

Alibaba (Qwen)General purposeOpen weights

Alibaba's first Qwen3.5 open model: a 397B MoE with 17B active under Apache 2.0, now outpaced by Qwen3.8 and its own hosted twin.

Sakana Namazu

Sakana AIGeneral purpose

Sakana AI's Japanese-tuned model built on Kimi K2.6, with built-in web search and code execution behind an OpenAI-style API.

Underdog Saluki 27B 1.0

Conway Research (Underdog)Small & on-deviceOpen weights

Qwen3.8-27B squeezed into a 7.89 GB 2-bit file that keeps tool calling intact, under Apache 2.0.

Xing4.0-29B-A4B

China Telecom (XingChen)General purposeOpen weights

China Telecom's Apache 2.0 agent model, notable less for its scores than for being trained entirely on Huawei Ascend chips.

HunyuanImage 3.0 Instruct

Tencent (Hunyuan)Image editingOpen weights

Tencent's 80B open image model with a reasoning step: third among open editors, but it needs an eight-GPU node and skips Europe.

Qwen3-TTS

Alibaba (Qwen)Text to speechOpen weights

Alibaba's Apache-licensed speech family: voice design from a sentence, cloning from three seconds, and no misuse safeguards.

SenseNova-Loop-9B-Preview

SenseTime (SenseNova)Vision & multimodalOpen weights

SenseTime's 9B image-and-text model that runs its middle 24 layers twice, aimed at office agents, under Apache 2.0.

SmolLM3-3B

Hugging FaceSmall & on-deviceOpen weights

Hugging Face's fully open 3B model, with switchable reasoning, a 128K context and its whole training recipe published.

all-MiniLM-L6-v2

Sentence Transformers (Hugging Face)Embeddings & rerankingOpen weights

The default embedding model of a thousand tutorials, now best kept for tests, prototypes and short English text.

Decision-2.0-Nox-4B (vLLM Semantic Router)

vLLM Semantic RouterDecision modelsOpen weights

vLLM Semantic Router's Apache-licensed 4B text decision model: well calibrated, but already outscored by the team's own d3-mini.

Devstral Small 2

Mistral AICodingOpen weights

Mistral's 24B Apache 2.0 coding model that runs on one RTX 4090 or a 32GB Mac, with a 256K context and image input.

KittenTTS 2

KittenML (Stellon Labs)Text to speechOpen weights

A 1.7B ternary speech model with voice cloning that downloads in under 1 GB and runs without a GPU, under a tight community licence.

Laya

Convai InnovationsDecision modelsOpen weights

Apache 2.0 decision model on ModernBERT that answers typed questions in tens of milliseconds, but needs fine-tuning to be accurate.

Phi-4-Reasoning-Vision-15B

MicrosoftReasoningOpen weights