Chat & reasoning

Run frontier and everyday LLMs locally with full control over sampling and reasoning.

Chat is the heart of the workbench. Run everything from a fast 7B to DeepSeek V4 Flash, fully on-device.

Screenshot: the chat view with the token inspector open — added at launch.

Models

Llama, Qwen, Gemma, Mistral, MiniMax M2, and DeepSeek V4 Flash (Q4Q8 / Q9). MoE models expose expert-count control where it’s safe; reasoning models expose effort modes (non-think / high / max).

Controls

Per chat or per model you can set context length, output budget, temperature, top-p, top-k, min-p, and penalties; toggle reasoning display; and turn on speculative decoding with a draft model where supported. Settings you change are previewed as “effective settings” merged with your global defaults.

Token inspector

See tokens, timing (time-to-first-token, tokens/sec), and — where the model supports it — logprobs, so you understand exactly what the model did.

Tools & structured output

Chat supports OpenAI-style tool calling and schema-constrained JSON. These each have their own page: Tool calling & MCP and Structured output, with API examples in API: OpenAI /v1.