Skip to content

Chat models

Open-weight Qwen and DeepSeek chat models on one OpenAI-compatible route.

Every chat model answers on POST /v1/chat/completions; model names one. GET /v1/models lists each model with its current state.

  • DeepSeek V4.1 Flashdeepseek-v4.1-flashMixture-of-experts model; 8,192-token context for prompt, reasoning and answer together, one request at a time.
  • Qwen3.8 27Bqwen3.8-27b150,000-token context, 64 parallel requests per account, tools and JSON output.
  • Qwen3.8 27B Fastqwen3.8-27b-fast65,536-token context, 8 parallel requests per account, tools and JSON output.
  • Qwen3.8 27B Uncensoredqwen3.8-27b-uncensoredRefusal behaviour removed; 150,000-token context, 64 parallel requests per account, tools and JSON output.
  • Qwen3.8 Flash-Next Abliteratedqwen3.8-flash-next-abliteratedMixture-of-experts model with refusal behaviour removed; 131,072-token context, one request at a time.

Capabilities

Context is in tokens, prompt plus output. Max parallel is per account (capabilities.max_parallel).

ModelIDContextMax parallelThinking switchTools and JSON output
DeepSeek V4.1 Flashdeepseek-v4.1-flash8,1921YesPassed to the model unchanged
Qwen3.8 27Bqwen3.8-27b150,00064YesYes
Qwen3.8 27B Fastqwen3.8-27b-fast65,5368YesYes
Qwen3.8 27B Uncensoredqwen3.8-27b-uncensored150,00064YesYes
Qwen3.8 Flash-Next Abliteratedqwen3.8-flash-next-abliterated131,0721YesPassed to the model unchanged
  • Thinking switch: chat_template_kwargs.enable_thinking: false turns the reasoning off; thinking is on by default (Thinking).
  • Tools and JSON output: The API sends tools, tool_choice and response_format to every model unchanged. On a model marked Yes, the reply carries tool calls in message.tool_calls and follows response_format (Tools and structured output).
  • Uncensored, Abliterated: In a model name, both words mean the same: the model's refusal behaviour was removed. Replies carry no safety warnings.