Qwen3.8-27B lands on Hugging Face: FP8 vision-language model with 1M-token context

Alibaba’s Qwen team has published Qwen3.8-27B-FP8, an FP8-quantized build of its newest 27-billion-parameter dense model. The release is a natively multimodal system that handles images and video alongside text, and it ships in Hugging Face Transformers format so it can be served locally through vLLM, SGLang, or Docker, or consumed via an OpenAI-compatible API. The quantization is fine-grained block-wise FP8 (block size 128), which the team says trims memory footprint while keeping accuracy essentially indistinguishable from the full-precision weights.

The more interesting details are architectural. Qwen3.8-27B uses a hybrid attention stack that mostly relies on linear-attention Gated DeltaNet blocks and inserts a full Gated Attention layer periodically across its 64 layers, and it adds multi-token prediction during training. Context length is 262K natively and is claimed to extend to one million tokens. The model also exposes explicit reasoning controls: a thinking mode that is on by default but can be switched off per request, a tunable reasoning_effort setting, and a preserve_thinking option that carries reasoning context across turns. The overall pitch is reliability on long-horizon, agentic coding and research work rather than raw one-shot answers.

Worth reading with a skeptical eye: every benchmark here is vendor-reported, and several were run through Anthropic’s Claude Code harness and compared against frontier closed models, with some scores drawn from in-house evals (QwenSWEBench, CoWorkBench). A hosted version on Qwen Cloud promising 1M context by default and built-in tools is described only as “coming soon.” Until independent evaluations appear, the leaderboard positioning and the near-lossless-quantization claim should be treated as marketing, not settled fact—but the open weights mean anyone can verify them directly.