Skip to content
LM15

Enter a model ID

GitHub
Playground

Tinker (Thinking Machines)

Tinker, by Thinking Machines, trains open models; it also answers chat requests for its base models and for any checkpoint you trained there. In LM15 it has two providers, one key:
Providers tinker (Chat Completions; also spelled thinkingmachines) and tinker-anthropic (Anthropic Messages)
Key TINKER_API_KEY, from tinker.thinkingmachines.ai/keys
Model names a base model, tinker:Qwen/Qwen3.5-4B or tinker:thinkingmachines/Inkling; a checkpoint you trained, by its path: tinker://<run>:train:0/sampler_weights/<name>
Needs the LM15 release after 2.2.0, in any language

Use tinker unless you have a reason not to: it lists your checkpoints, it can switch thinking off on every family that has an off switch, and it reports the cached tokens that make the cost exact. Tinker calls both doors beta. Measured against Tinker on November 9, 2026:

  • A checkpoint’s path is its name. tinker://… routes to Tinker as typed, colons and all. Listing the models of tinker shows your checkpoints and the base model each was trained from.
  • Thinking is on by default, and each model family switches it its own way; LM15 sends the right one. Inkling takes the effort levels. Qwen 3/3.5/3.6, Kimi K2.6 and Nemotron are an on/off switch: off turns it off, and any level turns it on (noted on the response). Qwen 3.8 and DeepSeek V4.1 take some levels and off. gpt-oss, Kimi K3.1 and GLM 5.3 can’t stop thinking; Tinker says so with an error.
  • A checkpoint thinks like the model it came from, but its path doesn’t say which that is. Tell LM15, and it sends that family’s switch: RouterConfig(base_models={"tinker://…": "Qwen/Qwen3.5-4B"}) in Python (baseModels in TypeScript, RouterConfig.BaseModels in Go, RouterConfig::base_models in Rust, RouterConfig(; base_models) in Julia, new_router(base_models = ) in R). Without it, off sends what most families take, which an Inkling checkpoint refuses.
  • A model you trained without thinking needs off to see the prompt it was trained on: Tinker builds the “thinking” prompt unless told not to.
  • No JSON mode. Tinker ignores response_format; LM15 leaves it out and notes that on the response. Describe the shape in the prompt.
  • Forcing a tool call is checked after the model answers: a model that answered without calling a tool comes back as an InvalidRequestError. Asking again may work.
  • Stop sequences work; Tinker also returns the sequence it stopped at, which LM15 removes, as on every other provider.
  • Images, in a message or in a tool result, reach the vision models (Inkling; Qwen3.5-4B refuses images). A PDF in a tool result is refused.
  • Cost is exact: cached prompt tokens are counted on both doors. A checkpoint is billed at its base model’s rates.
  • tinker-anthropic speaks the Anthropic Messages wire with the same key. It can’t switch thinking off on Qwen, Kimi K2.6 or Nemotron (Tinker refuses), and it has no model list.
  • A model or checkpoint that doesn’t exist raises UnsupportedModelError, as on every provider, with Tinker’s own message.