Is Qwen3.8-27B Still #1 for Local LLMs? (M5 Max Test)
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A hands-on first look at the newly released Qwen3.8-27B dense model, tested locally on an M5 Max Mac using LM Studio and the Pi coding agent. The model is compared against its predecessor Qwen3.6-27B and Meta's Muse Glimmer model on benchmarks like agentic terminal coding and repo-level code generation, where Qwen reportedly outperforms Muse Glimmer significantly (73 vs 51, 61.7 vs 51). The reviewer runs a 4-bit GGUF quantization (no MLX variant available yet), tests it on repo analysis, building an API feature, and generating a playable Mario Bros game in HTML/CSS/JS, reporting fast performance and decent tool-calling ability despite high memory usage (~40GB).
Questions this post answers
What is Qwen3.8-27B and how does it compare to Qwen3.6-27B and Meta's Muse Glimmer for coding?
Qwen3.8-27B is a newly released dense 27 billion parameter model, succeeding Qwen3.6-27B, designed for local coding agent use. On agentic terminal coding it scores 73 versus Muse Glimmer's 51, and on an agentic coding benchmark it scores 61.7 versus Muse Glimmer's 51, indicating significantly stronger coding performance. It ships initially as BF-16, with quantized GGUF versions (2-bit to 16-bit) available via Unsloth. Developers evaluating local coding models can track new Qwen releases and benchmarks on daily.dev.
Can I run Qwen3.8-27B locally on an Apple Silicon Mac with LM Studio?
Yes, the 4-bit GGUF quantization of Qwen3.8-27B runs on Apple Silicon machines like an M5 Max via LM Studio, though no MLX-optimized variant exists yet, only GGUF. Model storage is about 17GB, but actual memory usage with a larger context length can reach around 40GB, so sufficient unified memory is needed. Anyone sizing local LLM hardware for coding agents can follow real-world memory usage reports on daily.dev.