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

Alibaba's Qwen lab released Qwen 3.8 27B on Friday, a dense 27-billion-parameter multimodal model under the Apache 2.0 license ^[1][2], and it ships with `reasoning_effort` set to `xhigh` by default ^[3]. Anyone benchmarking it on a laptop this week will therefore measure a configuration nobody should run in production.

Start with sizing, because that is the part teams usually get right. The unquantized repository is 55.6GB before the inference runtime and KV cache ^[4]. Community MLX conversions for Apple silicon are about 16.1GB at 4-bit and 29.5GB at 8-bit, which makes a 32GB Mac a reasonable platform for the 4-bit build at moderate context ^[5]. Simon Willison ran a 17GB Q4_K_M GGUF in LM Studio on a 128GB M5 Max MacBook Pro and on an NVIDIA DGX Spark ^[6].

Now the part that will wreck your evaluation. Qwen's documentation lists three settings, `xhigh` (the default), `medium` and `low`, and the LM Studio GGUF preserves the `xhigh` default ^[3][7]. Willison first hit LM Studio's default 8,192-token context limit, because the model consumed all of it thinking about mundane prompts; loading the full 262,144-token context fixed that ^[8][9]. His pelican-on-a-bicycle SVG test then took 21 minutes, spending 22,276 reasoning tokens to emit 3,223 tokens of output ^[10]. That is 6.9 tokens of thinking per token delivered ^[11]. The same prompt with reasoning off produced 3,715 tokens in 137 seconds ^[12], roughly 9.2 times faster in wall clock and about 6.9 times cheaper in total tokens ^[13][14]. Decode speed is not the story: throughput only fell from about 27 tokens per second to about 20 as context grew, so the gap is volume ^[15]. Asked to "draw an svg of a circle", the model at `xhigh` spent several minutes deliberating over palettes and construction lines and returned an animated geometric study that was not what was asked for ^[16].

The quality is real. Willison calls it the best local pelican SVG he has generated, and still says the 21 minutes was not worth it; his recommendation is to start at `low` or with reasoning off ^[17][18][19]. Alibaba's own numbers put the model in the same league as Anthropic's Opus 4.6 at its Max setting ^[20] and show it beating the larger closed-weight Qwen 3.7-Plus on coding and office work ^[21][22], including a jump on DeepSWE from 14.2 to 42.2 ^[23], a gain of 28 points ^[24]. Those figures describe the original checkpoint, not the quantized builds most local users will actually load ^[25], and the harness matters as much as the model for agentic work ^[26]. Alibaba also reports leading Meta's Muse Glimmer-30B on every test where it publishes both scores ^[27].

Three things to watch. Independent benchmarks, which Willison notes have yet to weigh in ^[28]. Whether Alibaba ships a mixture-of-experts sibling in this size class, as it did for Qwen 3.6-35B, which would activate fewer parameters per token ^[29]. And the hosted Qwen Cloud version with one million tokens of context, extended via YaRN, which is where the long-context claims get tested without your RAM ^[30][31].