Find the right model by capability, cost, sovereignty and language.
Market map: price against capability
Every priced model, placed by blended price per million tokens and Valwood tier. Filled dots are open weights; rings are API-only. Click a dot for details, a region to isolate it.
Where models come from
Models per origin region, split by availability. Click a bar to explore that region.
Open-weight release cadence
Releases per month by origin, from Hugging Face publish dates.
Southeast Asian language scores
SEA-HELM scores (0–100) for every model in the registry that has been measured. Click a row for the model.
Context windows
How much text each model can take in at once. Click a bar to filter.
Language coverage
Models with a measured score or a declared language, for the 24 best-covered languages after English. Click a language to explore it.
Recommendations
Shortlist map
Fit score against monthly cost for every model that meets your constraints. Numbers mark the top six.
Have Claude review this shortlist against your goal and write a recommendation: primary pick, fallback, self-hosted option and what to test.
Compare
Up to three models side by side. Add models with the compare button on any card, or pick a preset.
Data & method
Where each number comes from, how the scores work, and what is covered.
Sources
Labs tracked
Method
Origin is where the developing lab is headquartered. It describes the publisher, not where a hosted request is processed: OpenRouter routes to third-party providers whose data locations vary. For hard residency requirements, self-host an open-weight model in your own region or contract a regional provider directly.
Tier (1–5) is a Valwood placement, not a benchmark score. It is based on the model's position in its family (flagship, mid, flash, mini), price band, parameter count and modalities. Validate finalists on your own evaluation set.
Blended price assumes three input tokens for every output token. Self-host memory is total parameters × 0.6 GB (4-bit weights plus headroom); mixture-of-experts models still need all weights in memory.
License comes from the Hugging Face license tag. "Custom" means terms such as Llama, Kimi, NVIDIA or Mistral research licenses that need a legal read before commercial use.
Language evidence has three levels. Measured: an independent leaderboard score. Declared: the model card or provider names the language. Claimed: a count such as "201 languages" without a list. No published data is common for closed API models and does not mean the model can't handle the language.
Recommendation score weighs task fit, quality match, cost at your volume, deployment and sovereignty fit, and language evidence when you name languages. The bar on each result shows how much each part contributed.
Next sources for measured language scores: the AI Language Proficiency Monitor (up to 200 languages), and regional boards such as Nejumi (Japanese), Open Ko-LLM (Korean), the Open Arabic LLM Leaderboard and EuroEval. To refresh the snapshot, ask Claude in the AI Model Registry project to re-pull the sources and republish this page.