Summary: Alibaba’s Qwen team put Qwen3.8-27B and FP8 weights on Hugging Face around August 13 under Apache 2.0, with roughly 27.8B parameters, native 262K context, and native image/video understanding. The model card reports 73.0 on Terminal Bench 2.1 and 84.3 on OSWorld-Verified, while Artificial Analysis independently scores it at 52. Simon Willison found it capable but slowed by the xhigh reasoning default, so teams should tune reasoning effort to the task rather than infer that its strongest results transfer to every workload.
Alibaba’s Qwen team wrapped up early August by opening the weights of its flagship Qwen3.8-Max, and in the same breath promised a smaller 27B sibling inside a week. That promise landed around August 13, when Qwen3.8-27B and an FP8 quantized build showed up on Hugging Face under the Apache 2.0 license, free for commercial use, with community GGUF quants ready on day one. The timing deserves a mention. Across that week, every closed-model headline was about subscription tiers and compute contracts, while the open-weights camp quietly nudged “capability you can run locally” one notch ahead. Only a month earlier, the worry had been that Qwen’s strongest models would stay locked behind subscriptions. With the 27B, a capable model under a permissive license is back on the download list.
The Specs: Compact, Long Context, Native Vision
Qwen3.8-27B carries roughly 27.8 billion parameters and a native context window of 262,144 tokens that YaRN can stretch to one million. Within the Qwen3.8 family it is the first small member to ship open weights, and it continues the Qwen3.5 foundation: 64 layers in a hybrid arrangement that alternates Gated DeltaNet and Gated Attention blocks, along with multi-token prediction (MTP). Nor is it a text-only model — image and video understanding are native. Thinking mode ships enabled, reasoning depth can be tuned with a reasoning_effort setting, and the official model card fixes the default at xhigh. That choice became the story. Benchmarks self-reported on the card put the model against Qwen3.6-27B, the closed-weight Qwen3.7-Plus, and a handful of proprietary rivals, claiming 73.0 on Terminal Bench 2.1 and 84.3 on OSWorld-Verified — figures that, as of this week, still lack independent verification. FP8 and community GGUF quantizations, vLLM, and SGLang support all existed from day one.
Willison’s Hands-On: A Capable Model Betrayed by Its Default
Simon Willison’s hands-on went up on August 16. He ran the 17GB Q4_K_M quantization in LM Studio on a 128GB M5 Max MacBook Pro, and also gave an NVIDIA DGX Spark a try. The verdict comes in two parts. On capability the model made a strong showing: bounding-box detection on photos is accurate enough that a single prompt produced a working labeling tool, and it steered the Pi coding agent through codebase questions plus a tested JSONL-to-Markdown converter. Reasoning genuinely earned its keep there — the variant without reasoning put boxes in the wrong places. On speed the default works against it: “draw a pelican on a bicycle” burned 22,276 reasoning tokens to emit 3,223 output tokens, across 21 minutes; with reasoning off, the same prompt wrapped up in 137 seconds. Even “draw an SVG of a circle” set off an elaborate animated design he never asked for. His advice is blunt: begin with low or no reasoning, and raise it only when a task deserves it.
Speed and Ecosystem
Local throughput landing between 15 and 30 tokens per second explains the rest. On Artificial Analysis’s leaderboards, hosted APIs currently deliver around 74 tokens per second for OpenAI 5.6 Sol and 184 for Luna — the dense-model tax on memory bandwidth, out in the open. Relief is partial: llama.cpp’s multi-token prediction draft decoding lifts local throughput by roughly 72%. Artificial Analysis has since added the model to its index with an independent score of 52, a helpful reality check on the self-reported figures. The ecosystem kept pace too: Unsloth GGUF quants, vLLM, and SGLang support on release day, and community patches are already out that flip the chat template default from xhigh to medium.
What It Means for Local Deployment
Three takeaways for teams. First, a single 17GB file now holds long context, tool calling, vision, and code generation — abilities that counted as flagship-class a year ago, running offline on a laptop today. That resets what on-device and air-gapped deployments can reasonably justify. Second, defaults are product. The xhigh setting turns a first impression into a long wait, the backlash came immediately, and the community is already shipping template edits that make medium the default. Vendors ought to tune defaults for the median task, not the hardest one. Third, the open-weights cadence shows no sign of slowing. Inside a month, Qwen open-sourced both its Max flagship and this 27B workhorse, delivering on its promise within days of the deadline. For teams that need private deployment — data that cannot leave the building, or per-token pricing that hurts — the local menu has gained another serious candidate, with independent benchmarks still pending.