Qwen3.8-27B puts frontier-class coding in a 17GB Apache 2.0 download

Alibaba's Qwen team has released Qwen3.8-27B, a dense, Apache 2.0, multimodal model with 262K native context that shrinks to roughly 17GB at 4-bit and carries vendor-claimed coding scores in frontier territory.

Key points

• Qwen/Qwen3.8-27B: Qwen's model card lists 27B dense weights under Apache 2.0, taking text, image and video input, with 262,144 tokens of native context stretchable to 1M. TNW puts the full-precision footprint around 56GB of GPU memory; a 4-bit build is near 17GB.
• The card describes a 64-layer hybrid stack that interleaves Gated DeltaNet linear-attention blocks with gated attention blocks at a 3:1 ratio, with 4 KV heads on the attention layers. Thinking is on by default, can be turned off per request, and a reasoning_effort setting offers xhigh (default), medium and low.
• Qwen's own reported scores, from the same card: 61.7 SWE-bench Pro, 73.0 Terminal-Bench 2.1, 90.3 LiveCodeBench v6, 89.2 GPQA Diamond, and 84.3 OSWorld-Verified on the vision side. All are the vendor's numbers.
• Independent figures via TNW, from August: Artificial Analysis rated it 52 on its Intelligence Index, level with GPT-5.6 Luna at maximum reasoning and top of a 135-model class. Artificial Analysis also gave it 51 on its separate Agentic Index, which TNW and VentureBeat both say beat Claude Opus 4.8 at maximum reasoning. Artificial Analysis has since revised the Intelligence Index; on version 4.3 its model page shows 34. TNW, citing Cybernews, also reports 3M Hugging Face downloads in the first three days.
• The cost is verbosity. Artificial Analysis logged 160M output tokens over its August test run versus a 43M median for similar open-weight models, and Simon Willison found a simple drawing request consumed 21 minutes and 22,000+ reasoning tokens at the default xhigh effort. Qwen's card names vLLM, SGLang and TokenSpeed for serving; the Hugging Face page also links community quantizations for llama.cpp, Ollama and LM Studio.
• Sibling Qwen/Qwen3.8-2.4T-A95B: a 2.4T-parameter MoE with 95B active, text-only, thinking always on, 262K native context. Its card reports 67.7 SWE-bench Pro and 86.6 Terminal-Bench 2.1 in a column labelled Qwen3.8-Max, the name of the hosted version built on this checkpoint with extra features such as vision input. Under its Qwen3.8-Max licence, a company running a Model-as-a-Service or AI Work Assistant business (the licence's term for standalone coding or office-productivity assistants) needs a separate licence from Qwen for commercial use once its group revenue tops US$50M in any 12 months, internal use excepted; products above 100M MAU or US$20M monthly revenue must also show the model name on screen.

“This is the first time a local model has scored frontier model capability.” — Cline, open-source coding tool, posting on X (as quoted by The Next Web)

Why it matters

For a two-person studio, a model Qwen says scores 61.7 on SWE-bench Pro, squeezed into about 17GB by community 4-bit builds, means a coding agent on a single high-end laptop or gaming PC with no per-token bill and no client code leaving the machine. Apache 2.0 settles the licence question for client work in a way the Max-class weights do not. Test latency before capability, though: independent testing found it burns three to four times the output tokens of comparable open-weight models, and it ships set to its highest reasoning effort, so an agent loop that looks free on paper may crawl in practice. Try reasoning_effort at medium or low before judging it.

This is an AI-written summary of the reporting credited above and the other sources linked in the text, read and edited by Nicholas before publishing. The facts belong to those sources and the quote to Cline, via The Next Web; the wording is ours. Read the original.