Qwen/Qwen3.6-27B · Hugging Face
4 hours ago
- #AI-Model
- #Multimodal-AI
- #Language-Model
- Qwen3.6-27B is a 27-billion-parameter causal language model with vision encoder, released as an open-weight variant prioritizing stability and real-world utility.
- Key upgrades include enhanced agentic coding for frontend workflows and repository-level reasoning, plus thinking preservation to retain reasoning context from historical messages.
- Model architecture features 64 layers, hidden dimension of 5120, and supports native context length of 262,144 tokens (extensible to ~1,010,000 via YaRN scaling).
- Benchmark results show strong performance in coding (e.g., SWE-bench Verified 77.2), language (MMLU-Pro 86.2), STEM (GPQA Diamond 87.8), and vision-language tasks (e.g., MMMU 82.9).
- Quickstart guides cover serving via SGLang, vLLM, KTransformers, or Hugging Face Transformers with OpenAI-compatible API, including setup for tool use and multi-token prediction.
- Usage examples demonstrate text, image, and video inputs via chat completions API, with sampling parameters recommended for thinking mode, precise coding, and instruct mode.
- Agentic capabilities are supported via Qwen-Agent and Qwen Code, enabling tool calling and automation.
- Best practices include setting adequate output length (e.g., 81,920 tokens for complex problems), standardizing output formats for benchmarks, and adjusting video processing for long videos.