Qwen 3.8 27B ships at 'xhigh' by default, and your latency budget pays for it

According to Qwen, Qwen3.8-27B outperforms the larger Qwen3.7-Plus in coding and office tasks, and the team also touts improved agent capabilities with more independent planning.

Alibaba reports a step up on the DeepSWE agentic coding benchmark from 14.2 points to 42.2 points versus the previous-generation Qwen 3.7-Plus.

The unquantized Qwen3.8-27B repository is 55.6GB, before accounting for the inference runtime and context cache.

Community-created MLX conversions for Apple silicon are available: the 4-bit version is about 16.1GB and the 8-bit version is 29.5GB, making a Mac with 32GB of unified memory a reasonable platform for the 4-bit version at moderate context length.

The model natively supports 262,000 tokens of context and can scale to one million using the YaRN method; a hosted version with one million tokens of context will soon be available through Qwen Cloud.

The full context window is not usable on a desktop because the key-value cache consumes more memory as the prompt grows.

Alibaba's benchmarks appear to describe the original checkpoint, not the quantized versions most local users will run, and quantization always carries a quality tradeoff.

The New Stack notes that benchmarks do not always measure real-world performance, that for agentic use cases the harness can matter as much as the model, and that early reports say the model tends to overthink.

Willison says it will be interesting to hear what independent benchmarks say about the model.

Willison ran the model on a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark, using LM Studio and a 17GB Q4_K_M quantized build, and also tried llama-server directly on the Spark.

The reported DeepSWE improvement is 28 points.

Willison ran into problems with LM Studio's default context limit of 8,192 tokens because the model used them all up thinking about even the most mundane problems.

Loading the model with the full 262,144 maximum context length made the context-exhaustion problem go away.

Willison's first pelican-riding-a-bicycle SVG attempt at the increased context length took 21 minutes to generate, using 22,276 reasoning tokens to produce 3,223 tokens of output.

The same pelican prompt run with reasoning turned off produced 3,715 tokens and took 137 seconds, just over two minutes.

Willison says the result is by far the best pelican SVG he has been able to generate with a model that runs on a local machine, noting the correct bicycle frame shape, legs on each side of the bike, a clear pelican pouch, wings touching the handlebars and motion lines behind rather than in front.

Asked whether the output was worth waiting 21 minutes for, Willison answers: absolutely not.

Willison's strong recommendation is to ignore the default and run Qwen 3.8 27B on low or even no reasoning levels at first.

According to Alibaba's benchmarks, Qwen 3.8 27B's performance is in the same league as Anthropic's Opus 4.6 running at its Max setting, and it outperforms that model on quite a few benchmarks, especially some computer use, coding and knowledge work tests.

Alibaba compares Qwen3.8-27B with Meta's Muse Glimmer-30B, a recently released local model in roughly the same size class, and reports Qwen leading on every test for which it publishes scores for both models.

Alibaba has not released a smaller Qwen3.8 27B mixture-of-experts sibling; such a model could activate fewer parameters per token and run faster than the dense 27B release, and Alibaba did release one for Qwen 3.6-35B.

At the xhigh default, the pelican run spent about 6.9 reasoning tokens for every token of output.

The xhigh run took about 9.2 times longer in wall clock than the same prompt with reasoning off.

The xhigh run generated about 6.9 times as many total tokens as the reasoning-off run.

Average token throughput was about 20 tokens per second on the xhigh run versus about 27 tokens per second with reasoning off, so the wall-clock difference is driven by token volume rather than a large drop in generation speed.