最新模型
Qwen3.8-27B Delivers a Native 262K Context Window with a Hybrid Linear/Full-Attention Architecture While Retaining Multi-Step MTP
Qwen has released the weights for Qwen3.8-27B, using a 64-layer hybrid Gated DeltaNet/Gated Attention architecture to reduce long-context costs while providing native visual understanding and adjustable reasoning depth. Official agent benchmarks show substantial gains, but some datasets, corrected tasks, and execution frameworks were configured by Qwen, so the results should not be treated as a fair cross-vendor ranking.
Qwen3.8-27B is a 27-billion-parameter dense model licensed under Apache 2.0 that handles text, images, and video. Its language backbone has 64 layers arranged as 16 groups of “three Gated DeltaNet linear-attention layers followed by one Gated Attention layer.” The former uses 48 value heads and 16 QK heads, while the latter uses 24 query heads and 4 KV heads. The key benefit of this hybrid design is that it avoids making every token pay the full quadratic attention cost at every layer, while periodically using full attention to exchange global information.
The model has a native context length of 262,144 tokens, which Qwen says can be extended to 1 million tokens. Training also incorporates multi-step Multi-Token Prediction, allowing compatible runtimes to use MTP for speculative decoding. Reasoning mode is enabled by default. Through the API, reasoning_effort can be set to xhigh, medium, or low, while preserve_thinking carries reasoning state from earlier turns into subsequent messages. This changes how agent integrations should be designed: if a chat gateway strips historical reasoning, results may differ from the official configuration. Lower reasoning effort may also fail to reduce total task cost, because retries after failures can offset the tokens saved on each attempt.
Qwen officially reports scores of 73.0 on Terminal-Bench 2.1, 61.7 on SWE-bench Pro, and 84.3 on OSWorld-Verified, all above the listed results for Qwen3.6-27B. However, SWE-bench Pro uses a Claude Code harness and includes some corrected tasks. QwenSWEBench, CoWorkBench, and RecreationBench are also internal evaluations, while some visual tasks use different prompts or computer-interaction settings. Engineering teams should rerun evaluations with their own toolsets, context-pressure conditions, and error-recovery workflows before deciding whether to replace an existing model. Transformers, vLLM, and SGLang already provide loading paths, and the community has quickly produced GGUF, MLX, and FP8 variants. The next thing to watch is whether each runtime can correctly support hybrid-state caching, MTP, video sampling, and reasoning-state preservation across long conversations—not merely load the weights successfully.