Open-weight LLM landscape

A Alibaba Z Zhipu AI M Moonshot AI D DeepSeek M MiniMax (minimax.io) T Tencent T Thinking Machines X Xiaomi N Nvidia M Meta M Mistral G Google O OpenAI M Mistral C Cohere A Arcee AI

Blended cost = (7·cache + 2·input + 1·output)/10. When a provider omits cache-hit pricing, the input price is used as an upper bound (cache hits are never more expensive than a regular input token). 47 models shown: open-weight (via models.dev), with an Agentic Index score AND input + output pricing. Models without input or output pricing are excluded. Energy per request (16k–64k band) is measured (solid dots) for 7 models with a Neuralwatt match, and estimated as k × blended cost (hollow dots) for the remaining 36 — see the Neuralwatt calibration below.

Calibration view: measured energy per request (mWh, 16k–64k band, scraped from the Neuralwatt portal) against NW blended cost per 1M tokens = (7·cache + 2·input + 1·output)/10 (USD). Dashed line: proportional fit energy = k × cost through the origin, fitted on base models only (excluding -fast / -short variants); this k is used to estimate energy for the models without a Neuralwatt match in the comparison chart above. Hover any dot for details.