A 4-bit Mac build fits Qwen3.8-27B in 16.1GB

Alibaba on August 14 published the weights of Qwen3.8-27B, a dense 27-billion-parameter model, under Apache 2.0. Alibaba’s own benchmarks put it in the same league as Anthropic’s Opus 4.6 running at its Max setting, The New Stack reports, and the model is small enough to run on a well-specced MacBook Pro or Mac Studio.
The 27B model came after a much bigger sibling. Two days earlier, on August 12, Alibaba published Qwen3.8-2.4T-A95B, the 2.4-trillion-parameter version, under a bespoke Qwen3.8-Max License with revenue-share conditions, and that model is text-only. The 27B checkpoint ships under plain Apache 2.0 and handles images and video. So the smaller release is the more open of the two, and the only one that can see. Alibaba describes Qwen3.8 as the first time a Qwen-Max-class model has had an open release, built on the Qwen3.5 architecture.
On quite a few benchmarks, Alibaba’s charts put the 27B model ahead of Opus 4.6, which was state of the art when it launched in February, with the biggest margins in computer use, coding and knowledge work tests. Against the previous-generation Qwen 3.7-Plus, the DeepSWE agentic coding score went from 14.2 to 42.2. That still trails the current frontier and other open models such as the newly released GLM-5.3, though Google’s mid-tier Gemini 3.6 Flash only reached 49% on the same test, and you are not running Flash on a laptop anytime soon. Alibaba also lines it up against Meta’s Muse Glimmer-30B, a local model in roughly the same size class that needed more than 55 GB at full precision when Meta released it in August. Qwen3.8-27B leads Glimmer on every test where Alibaba reports scores for both.
Those scores come with the usual caveats, and one of them applies directly to local use. Alibaba’s benchmarks appear to describe the original checkpoint, not the quantized builds most people will actually download, and quantization always gives up some quality. For agentic work, the harness the model runs in can matter as much as the model. Early reports also say the model tends to overthink. Alibaba’s Qwen3.8 repository lists a reasoning_effort setting that tunes reasoning depth, and preserve_thinking, which keeps reasoning from earlier messages in the conversation history.
Speed is the other open issue, and it follows from the model being dense. A mixture-of-experts model (one that sends each token through only a subset of its weights) can activate fewer parameters per token and run faster. Alibaba has not released such a variant of Qwen3.8 at this size, though it did ship one in the Qwen 3.6 generation, the Qwen 3.6-35B model.
The unquantized repository is 55.6GB, and that is before the inference runtime and the context cache. Few people will run it that way. Community-made MLX conversions for Apple silicon are already up: the 4-bit build is about 16.1GB and the 8-bit build is 29.5GB. That makes a Mac with 32GB of unified memory a reasonable platform for the 4-bit version at a moderate context length, while 48GB or 64GB gives considerably more room for higher precision or longer prompts.
The model supports 262,000 tokens of context natively. With YaRN, a technique that stretches a model’s position handling past the length it was trained on, that goes up to 1 million tokens, and Alibaba will turn that on for its hosted production version. A desktop is a different matter: the key-value cache, the stored attention state for every token already in the prompt, consumes more memory as the prompt grows, so the full window is not something a local machine automatically gets to use.
Up next
Why output tokens in Claude Code cost about 5x what input does
Switching models or changing your /effort level mid-conversation in Claude Code wipes the prompt cache, and the entire thread gets re-counted at full input price, per Anthropic. Cache reads run 0.1x input; the cache lasts an hour on a subscription, five minutes on an API key.
Related stories
- 27B open weights ship under Apache 2.0
- Mistral Large 4 weights are out, API now from $0.70 per 1M
- Kimi K3.1 Pro lands just behind Opus 5.5 on Terminal-Bench
- DeepSeek’s minor Flash update beats Opus 5 on DeepSWE
- Liquid AI’s 2B-active model beats Qwen on LiveClaw
- Mistral Large 4 tops open models at just 15% on legal tasks
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

