DeepInfra

DeepInfra runs open models from many vendors (DeepSeek, Qwen, GLM, Kimi, Llama, gpt-oss and others) behind one API. In LM15 its provider is deepinfra.

Provider	deepinfra (litellm spelling: deepinfra/)
Key	DEEPINFRA_API_KEY, from deepinfra.com/dash/api_keys
Model names	the vendor’s name: deepinfra:deepseek-ai/DeepSeek-V4.1-Flash
Needs	Python 1.1.0, TypeScript 1.0.0-rc.2, Rust 1.0.0-rc.2, Go v1.1.0-rc.2 or later

Chat, streaming, tools, structured output and the model list work as with any provider. What LM15 does differently here, each measured against DeepInfra on September 26, 2026:

• Forcing a tool call. Many models on DeepInfra ignore a forced tool choice (required, a named tool, or none) and answer in text. LM15 sends it only to the 14 models measured to honour it: DeepSeek V3.2, V4-Flash and V4.1-Flash, GLM-5.3-Flash, Kimi K2.6, Llama 4 Scout, Qwen3.6-27B, Qwen3-Next, Nemotron 3.5 Lightning, Granite 4.2 8B, MiMo V2.6 Flash, Hunyuan 3, Gemini 3.1 Flash Lite and Claude Haiku 4.5. On any other model it raises UnsupportedFeatureError before sending anything, instead of letting the request be ignored.
• Reasoning off on gpt-oss. gpt-oss can’t stop reasoning, and DeepInfra accepts “off” and runs it at a low level anyway. LM15 sends the lowest level itself and says so in the response’s adaptations.
• Images inside a tool result are refused: DeepInfra accepts only text there.
• Caching is automatic. A cache key or long retention has no place in the request LM15 sends, so it is left out and noted in the adaptations.