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

Alibaba released a dense 27 billion parameter version of Qwen3.8 under the Apache 2.0 license on Friday, alongside the open weights of the 2.4 trillion parameter flagship model, in a build small enough to run on a well-specced MacBook Pro or Mac Studio, according to The New Stack.

At a glance

• The unquantized repository weighs 55.6GB, while community MLX conversions for Apple silicon land at 16.1GB in 4-bit form and 29.5GB in 8-bit, before inference runtime and context cache.
• On the DeepSWE agentic coding benchmark the model scores 42.2 points against 14.2 for the previous-generation Qwen 3.7-Plus, still short of Google's mid-tier Gemini 3.6 Flash at 49%.
• The New Stack reports that Alibaba's benchmark numbers describe the original checkpoint rather than the quantized builds most local users will run, and that early reports say the model tends to overthink.

Why it matters The interesting number here is not the leaderboard position but the memory footprint. Weights that fit in 16.1GB put frontier-adjacent capability on hardware a large share of developers already own, which likely shifts the calculus for work that cannot leave a local machine. Benchmark suites measure something narrower than daily use, though, and for agentic workloads the harness around a model appears to matter as much as the weights themselves.

Alibaba's benchmarks place the 27B model level with Opus 4.6 at its Max setting

According to Alibaba's own numbers, the dense 27 billion parameter Qwen3.8 performs in the same league as Anthropic's Opus 4.6 running at its Max setting, and comes out ahead on quite a few tests, particularly computer use, coding and knowledge work. Opus 4.6 was state of the art when it launched in February.

The 27B release accompanied the open weights of the 2.4 trillion parameter Qwen3.8, whose benchmarks put it in direct competition with closed frontier models from American labs. The dense 27B model also has vision capabilities, including video, which The New Stack notes makes it a compelling option for local use.

The DeepSWE score climbs from 14.2 to 42.2 points

On DeepSWE, an agentic coding benchmark, Qwen3.8-27B reaches 42.2 points against 14.2 for Qwen 3.7-Plus. That still leaves it behind the current frontier and behind other open models, including the recently released GLM-5.3. Google's mid-tier Gemini 3.6 Flash reached 49% on the same test.

Alibaba says the model significantly outperforms the previous-generation Qwen 3.7-Plus, especially on coding and knowledge work tasks, and compares it against Meta's Muse Glimmer-30B, another recently released local model in roughly the same size class, reporting a lead on every test where scores for both appear.

Benchmarks do not always track real-world performance, and for agentic use cases the harness a model runs in can matter as much as the model itself. Alibaba's figures describe the original checkpoint rather than the quantized versions most local users will install, and quantization always carries a quality tradeoff.

A 4-bit MLX build fits in 16.1GB against 55.6GB unquantized

The unquantized Qwen3.8-27B repository is 55.6GB before the inference runtime and the context cache are accounted for, which is not what most people will run. Community MLX conversions for Apple silicon are already available: the 4-bit version is about 16.1GB, the 8-bit version 29.5GB.

That makes a Mac with 32GB of unified memory a reasonable platform for the 4-bit build at moderate context length, while 48GB or 64GB leaves considerably more room for higher precision or longer prompts. The model supports 262,000 tokens of context by default, extendable to 1 million tokens through the YaRN method. The key-value cache consumes more memory as the prompt grows.

What's next

Alibaba has not released a smaller Qwen3.8 mixture-of-experts sibling, which would activate fewer parameters per token and run faster than the dense 27B build. It previously released such a model for Qwen 3.6-35B. Alibaba will extend the context window to 1 million tokens with YaRN for its hosted production version of Qwen3.8.