In early August, when Alibaba’s Qwen team released open weights for its flagship Qwen3.8-Max, it promised that a smaller 27B sibling would follow within a week. Around August 13, the promise landed: Qwen3.8-27B and an FP8 quantized version appeared on Hugging Face under the Apache 2.0 license, free for commercial use, with community GGUF quants ready the same day. The timing is worth noting. In the same week, the closed-model news cycle revolved around subscription tiers and compute contracts; the open-weights side quietly moved “capability that runs locally” one notch forward. A month ago the worry was that Qwen’s best models would stay locked behind subscriptions. The 27B puts a capable, permissively licensed model back on the download list.
The Specs: Small, Long Context, Native Vision
Qwen3.8-27B has roughly 27.8 billion parameters, a native context window of 262,144 tokens extendable to one million via YaRN. It is the first small member of the Qwen3.8 family to ship open weights, built on the Qwen3.5 architecture: 64 layers in a hybrid layout alternating Gated DeltaNet and Gated Attention blocks, plus multi-token prediction (MTP). It is not a text-only model either — image and video understanding are native. Thinking mode is on by default, reasoning depth is tunable through a reasoning_effort setting, and the official model card sets the default to xhigh. That choice became the story. The card’s self-reported benchmarks pit it against Qwen3.6-27B, the closed-weight Qwen3.7-Plus, and several proprietary rivals, claiming 73.0 on Terminal Bench 2.1 and 84.3 on OSWorld-Verified — numbers that, as of this week, still lack independent verification. FP8 and community GGUF quantizations, vLLM, and SGLang support were available from day one.
Willison’s Hands-On: Capable Model, Costly Default
Simon Willison published his hands-on on August 16. He ran the 17GB Q4_K_M quantization in LM Studio on a 128GB M5 Max MacBook Pro and also tried an NVIDIA DGX Spark. The verdict splits in two. On capability, the model impressed: bounding-box detection on photos is accurate enough that a usable labeling tool emerged from a single prompt, and the model drove the Pi coding agent through codebase questions and a tested JSONL-to-Markdown converter. Reasoning genuinely helped there — the no-reasoning variant misplaced boxes. On speed, the default sabotages it: “draw a pelican on a bicycle” burned 22,276 reasoning tokens to emit 3,223 output tokens and took 21 minutes; with reasoning off, the same prompt finished in 137 seconds. Even “draw an SVG of a circle” triggered an elaborate animated design he never asked for. His advice is blunt: start with low or no reasoning and turn it up only when the task earns it.
Speed and Ecosystem
Local throughput of 15 to 30 tokens per second tells the rest of the story. Hosted APIs currently serve around 74 tokens per second for OpenAI 5.6 Sol and 184 for Luna on Artificial Analysis’s leaderboards — the dense-model tax on memory bandwidth, made visible. There is partial relief: llama.cpp’s multi-token prediction draft decoding pushes local throughput up by roughly 72%. Artificial Analysis has since indexed the model with an independent score of 52, a useful sanity check against the self-reported numbers. The ecosystem moved fast: Unsloth GGUF quants, vLLM, and SGLang support on release day, and community patches already exist that flip the chat template default from xhigh to medium.
What It Means for Local Deployment
Three takeaways for teams. First, a 17GB file now contains long context, tool calling, vision, and code generation — flagship-class ability a year ago, now running offline on a laptop. That resets expectations for what on-device and air-gapped deployments can justify. Second, defaults are product. The xhigh setting turns a first impression into a long wait, the backlash was immediate, and the community is already shipping template edits that make medium the default. Model vendors should tune defaults to the median task, not the hardest one. Third, the open-weights cadence is not slowing down. Within a month, Qwen open-sourced both its Max flagship and this 27B workhorse, following through on its promise within days of the deadline. For teams that need private deployment — data that cannot leave the building, or cost curves that make per-token pricing painful — the local menu just gained another serious candidate, with independent benchmarks still to come.
Sources
- Qwen3.8-27B-FP8 model card — Hugging Face
- Qwen 3.8 27B is excellent, but it defaults to overthinking things — Simon Willison
- Qwen3.8 collection — Hugging Face
- Qwen3.8 27B — Artificial Analysis
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.