What is the Model Atlas?
A reference, not a leaderboard. It lists the model families a practitioner would actually choose between, from the frontier labs down to sovereign programs with no published score at all, and for each one records what a leaderboard leaves out: the license you would be shipping under, the price per million tokens, the context window, what the model takes in, what hardware it fits on, and what it was built on top of. Embedding, reranking and speech models get the same treatment in their own sections.
Every figure comes from the vendor’s own page, its model card, or a primary file. Where a vendor publishes no figure, the cell says so rather than carrying a reseller’s estimate, and where a figure is derived rather than reported, the method is printed beside it.
It carries a state of record date rather than pretending to be permanent, and it is rebuilt on a schedule, with a changelog so anyone holding the last edition can see what moved. The PDF is free and there is no form.
Thirty-six of the 53 downloadable models here are plain Apache 2.0 or MIT. Of the other 17, 16 carry conditions that bite at a named revenue threshold and 1 states no terms you can pin down. Huawei's openPangu-2.0-Pro forbids use in the European Union.
This is an AI model license comparison across 70 model families from 14 countries. Each row carries who built the model, whether you can download the weights, what it costs per million tokens, how it scores, and what it is honestly good and bad at.
Leaderboards rank models and pricing pages price them. Neither answers the two questions that arrive after deployment, which are whether you are allowed to put the model in front of a client and what it was built on top of. Both get a column here.
This page carries the shape of the market: the license split, the four constraints people actually buy against, the top of the board, and where the 70 families sit by origin. The row-by-row record, every model with its license, price, context window and provenance, is in the PDF.
Every figure comes from the vendor's own page, its model card, or a primary file, and the method behind that is at the bottom. For what to check before you run one of these, see the open weight checklist. For where the API bill is heading, see AI token price increases.
Take the Atlas with you
Every one of the 70 families, eight fields against each, laid out for printing and for sending to a client.
- All 70 families in one table, grouped by where each one was built
- License, price in and out, context window, parameter count, input modality and hardware fit on every row
- 26 embedding, reranker and speech families on the same terms
- What each model is good at, what it is bad at, and what it was built on top of
- The method, the caveats and the known gaps, written down rather than implied
- Refreshed monthly!
30 page PDF // No Form or Email Required
FAMILIES LISTED
Across 14 countries of origin
OPEN WEIGHTS
Downloadable and self-hostable
BEST OPEN MODEL
MiMo-V2.6-Pro, Xiaomi
BEST US OPEN MODEL
Inkling-Small, Thinking Machines
Start with the license
36
Plain Apache 2.0 or MIT
16
Conditional Terms
1 - Terms
Not Stated
Conditional terms bite at real thresholds: Kimi K3 needs a separate agreement with Moonshot above $20M revenue; Mistral Medium 3.5 withdraws rights above $20M monthly revenue; openPangu-2.0-Pro forbids use in the European Union. The one unstated is Ling-3.0-Flash, which declares MIT with no license file. Open weights is not open source: of the 53, two publish the data, the checkpoints and the code.
Then pick by the constraint you are actually under
Lowest cost per finished task
License certainty for a client contract
Runs on hardware you already own
You must show an auditor what went in
Where the 70 families sit
By where each one was built and whether its weights are public.
13
US
Closed
14
US
Open
18
China
Open
21
Elsewhere
Open
4
Other
Close
Thirteen of the American models are closed, and only twelve of those publish an index score at all. The fourteen American open-weight models top out at 26. The eighteen Chinese ones hold five of the top five open scores, starting at 46, which puts 20 points between the best open model anywhere and the best American one.
The twenty-one from everywhere else are the harder problem. They come from twelve countries, and nine of them have no published index score at all. The best of those that do, Motif-3 and K2 Horizon, score in the low thirties. Sovereign AI programs are shipping weights faster than anyone is benchmarking them.
For reference, the top of the board
Artificial Analysis Intelligence Index v4.3, highest published reasoning-effort setting. A tier, not a ranking.
MiMo-V2.6-Pro
46
GLM-5.3
45
Kimi K3
44
GLM-5.3-Flash
42
Inkling-Small
26
Claude Opus 5.5
58
Beyond text generation
Twenty-six embedding, reranker and speech families, listed on the same terms. The score column is the public board for that section: MTEB Multilingual v2 for embeddings and rerankers, and Open ASR Leaderboard average word error rate for speech, where lower is better. These rows are not in the cover counts or the licence ladder.
Embeddings, 9 listed: Harrier-OSS-v1 27B (Microsoft (Bing), MIT, MTEB 74.3); Qwen3-Embedding (Alibaba, Apache 2.0, MTEB 70.6); EmbeddingGemma 2 (Google DeepMind, Apache 2.0, MTEB 61.4); Gemini Embedding 2 (Google DeepMind, API); text-embedding-3-large (OpenAI, API); Embed 5 (Cohere, API); voyage-4-large (Voyage AI (MongoDB), API); jina-embeddings-v5 (Jina AI, CC BY-NC 4.0, MTEB 67.0); BGE-M3 (BAAI, MIT).
Rerankers, 8 listed: Qwen3-Reranker (Alibaba, Apache 2.0); Rerank 4 (Cohere, API); rerank-3 (Voyage AI (MongoDB), API); jina-reranker-v3.5 (Jina AI, CC BY-NC 4.0, MTEB 67.8); bge-reranker-v2-m3 (BAAI, Apache 2.0); zerank-2 (ZeroEntropy (Notion), Apache 2.0); mxbai-rerank-large-v2 (Mixedbread, Apache 2.0); llama-nemotron-rerank-1b-v2 (NVIDIA, OpenMDW-1.1).
Speech: recognition and synthesis, 9 listed: Whisper large-v3 (OpenAI, Apache 2.0, WER 5.78); GPT-Transcribe, GPT-Realtime 2.1 (OpenAI, API); Parakeet TDT 0.6B v3, Canary-Qwen 2.5B (NVIDIA, CC BY 4.0, WER 4.86); Qwen3-ASR, Qwen3-TTS (Alibaba, Apache 2.0, WER 4.31); Voxtral (Mistral AI, Apache 2.0; TTS CC BY-NC 4.0, WER 4.99); Scribe v2, Eleven v4 (ElevenLabs, API, WER 3.97); Kyutai STT, TTS, Pocket TTS (Kyutai, CC BY 4.0, WER 5.57); Nova-3, Flux, Aura-2 (Deepgram, API); Saaras V4, Bulbul v3 (Sarvam AI, API).
What the AI model license comparison covers
Every family gets ten fields in the PDF: what it is called and who built it, where it was built, whether the weights are downloadable and under what license, parameter count, context window, price in and out, index score, what it takes in, what it fits on, and three lines on what it is good at, what it is bad at, and what it was built on top of.
Prices are USD per million tokens, input then output, as each vendor quotes them. A blank price is deliberate and means the vendor publishes no list price, which is not the same as a low one.
That last field is the one nobody else publishes. Japan's largest homegrown model ships a config file naming DeepSeek's architecture, and the model card does not mention it. Provenance is the column that tells you whose problem it becomes when the upstream model is disputed.
Take the Atlas with you
Every one of the 70 families, eight fields against each, laid out for printing and for sending to a client.
- All 70 families in one table, grouped by where each one was built
- License, price in and out, context window, parameter count, input modality and hardware fit on every row
- 26 embedding, reranker and speech families on the same terms
- What each model is good at, what it is bad at, and what it was built on top of
- The method, the caveats and the known gaps, written down rather than implied
- Refreshed monthly!
30 page PDF // No Form or Email Required
How to read this, and what it does not contain
What verified means here. Every figure came from the vendor's own page, its model card, or a primary file. Where a vendor publishes no figure, the cell says so rather than carrying a reseller's number. Two claims were checked against primary files rather than press coverage: Rakuten AI 3.0's config.json, and the leaderboard scores.
There are no unsourced parameter counts for closed models here, and no prices from aggregators.
Every row in the PDF carries two columns that need a key: what the model takes in, and what it fits on.
What Fits means
A band derived from total parameters, not a vendor claim. It assumes 4-bit weights, about 0.55 bytes per parameter plus headroom for the KV cache, and says what you can load, not what runs fast.
Tiny. Under about 12B. A desktop GPU with 8GB to 12GB, a Jetson Orin Nano, a phone, or a Raspberry Pi 5 at the low end, where only the 4B-and-under sizes are practical.
Small. 12B to about 35B. One consumer GPU of 16GB to 24GB, a 32GB M4 Mac mini, or a Jetson Orin.
Medium. 35B to about 200B. One 48GB to 80GB datacenter GPU, or a 64GB to 128GB unified-memory machine: an M4 Pro Mac mini at 64GB, a Mac Studio, a Jetson Thor or a DGX Spark at 128GB.
Large. 200B to about 1.2T. A node of two to eight datacenter GPUs, 2x to 8x H100, H200 or B200.
Datacenter. Above about 1.2T. A node of 288GB-class GPUs or more than one node.
API. Closed weights. Nothing to buy; the vendor runs it.
Total parameters set memory even when few are active: a 125B model with 6B active still needs about 70GB. A family that ships several sizes shows a range, Tiny to Small. Where a vendor states a minimum or documents its smallest deployment, it is printed under the band; those figures are usually at BF16 or FP8 and so heavier than the band.
What Takes in means
Input modality as the model card states it. Output is text unless a row says otherwise; embeddings return a vector and speech rows list both directions.
text. Text only. The card or config declares no image, audio or video encoder.
text, image. Images in: screenshots, documents, photos. Video, if accepted at all, as separate frames.
text, image, video. Video as a sequence of frames, by the card's own description.
text, image, audio. Speech or sound in as audio, not transcribed first.
text, image, audio, video. All four. Only Gemini, Muse Spark, MiMo-V2.6 and the omni embedders.
not stated. The vendor's page is silent. Two closed rows.
A modality the API accepts is not the same as one the open weights accept: Qwen3.8-Max's base download is text only while the Max API takes images.
- Everything has a shelf life of weeks. Re-check the vendor page before quoting a number.
- Version numbers stopped meaning anything. OpenAI shipped GPT-6 Sol and GPT-6.1 Sol a week apart and dropped the Terra tier. Google's Flash line is at 3.8 while Pro sits at 3.1 and is still labelled preview.
- Index rows are per reasoning-effort setting. Scores here are the highest published setting. Treat them as a tier, not a ranking.
- These are v4.3.2 scores. The index was rescored in September. Higher figures circulating from launch coverage, GLM-5.3 at 60 and Nemotron 3 Ultra at 47.7, are from the previous version.
- Gemini 3.8 Flash pricing is promotional. The $0.75 / $3.75 rate holds through 31 December 2026 and doubles on 1 January 2027.
- DeepSeek prices vary by time of day. Off-peak is half of peak. Off-peak rates are shown throughout. The V4-Flash endpoint was retired on 10 September.
- Blank price cells are deliberate. NVIDIA, Gemma, gpt-oss, IBM, Cohere and most sovereign programmes sell no API and so have no list price. A blank cell is not a low number.
- Open weights is not open source. Open weights means you can run the parameters. Open source means you can rebuild the thing.
Known gaps, carried rather than hidden.
Qwen3.8-Max. The downloadable weights are the Qwen3.8-2.4T-A95B base, feature-reduced against the Max API product. Treat the row as the base model.
Ling-3.0-Flash. MIT is declared in the repository metadata and there is no license file. The other three September license questions were resolved from their repositories.
SEA-LION v4.5. AI Singapore declares MIT on both series, but the Gemma-based one links Google's own license page. Which terms bind you is for AI Singapore to clarify.
Africa. No significant 2025 or 2026 African foundation-model release was found. That is a research gap rather than evidence of absence.
Announced, not shipped. Reflection AI's Beam (501B, Apache 2.0 stated, announced 5 October) and StepFun's Step 5 Preview have weights promised but not public. Neither is a row until the weights are downloadable.
Sources: vendor pricing pages and model cards, Hugging Face model repositories and its “State of Open Models, Summer 2026” analysis, and the Artificial Analysis leaderboard read live. Refreshed monthly.
Signup to get updates by email!
There's nothing I hate more than spam, except perhaps painting, so fear not by subscribing.
