Skip to content
NewsModel release

Alibaba releases Qwen3.8-27B with a 3x DeepSWE jump and 1M-token context on a single RTX 4090

· by Pondero Newsdesk

The short version

Alibaba shipped Qwen3.8-27B on August 14 under Apache 2.0. The 27.78B-parameter dense model scores 42.2 on DeepSWE 1.1 and 84.3 on OSWorld-Verified, both from a single 24GB VRAM GPU.

Alibaba releases Qwen3.8-27B with a 3x DeepSWE jump and 1M-token context on a single RTX 4090

Alibaba's Qwen team released Qwen3.8-27B on August 14 with a number that is hard to ignore: DeepSWE 1.1 agentic coding went from 13.3 to 42.2, a more-than-three-fold jump over the prior 27B generation, in a model that fits on a single consumer GPU.

What

Qwen3.8-27B ships with Apache 2.0 weights and accepts text, images, and video as input. Native context is 262,144 tokens, extensible to one million via YaRN scaling. The hardware target is a single 24GB VRAM card (RTX 4090 class).

Benchmark gains over the predecessor Qwen3.6-27B are sharp across multiple categories, per the HuggingFace model card:

  • Agentic coding (DeepSWE 1.1): 13.3 to 42.2
  • Computer use (OSWorld-Verified): 63.9 to 84.3
  • Terminal coding (Terminal Bench 2.1): 63.4 to 73.0
  • Browser use (WebArena-Verified): 48.8 to 64.8
  • Mobile agent (AndroidWorld): 70.3 to 81.9

The model scores 61.7 on SWE-bench Pro and 90.3 on LiveCodeBench v6. Thinking mode is active by default. Operators can adjust reasoning depth via a reasoning_effort parameter (xhigh, medium, or low), which lets them trade compute cost against task complexity per request.

Why it matters

Agentic-coding and computer-use scores at this level were not achievable in an open-weights 27B model before this release. DeepSWE at 42.2 and OSWorld at 84.3 are figures that, until now, required either a much larger model or a proprietary API call. Both are in the range where production pipelines start becoming realistic, not just research experiments.

The competitive frame matters too. Meta's Muse Glimmer-30B, released earlier in August, posted 51.7 on Terminal Bench 2.1 and 65.9 on OSWorld-Verified per OfficeChai's benchmark breakdown. Qwen3.8-27B clears both marks. No Muse Glimmer result is reported on DeepSWE 1.1, so a direct agentic-coding comparison is unavailable, but on the benchmarks where both appear, the Alibaba model leads.

For a team selecting a self-hosted coding agent, the calculus shifted on August 14. Apache 2.0 removes the deployment restrictions that proprietary models carry. The single-GPU requirement keeps infrastructure costs low. Teams already running Ollama, vLLM, or SGLang can pull the weights and evaluate the model today without waiting for a hosted version.

What to watch next

Community benchmarks against Qwen3-30B-A3B, the mixture-of-experts variant from the same generation, should appear within the week. Apple Silicon MPS optimization patches are expected within one to two weeks, extending the consumer-hardware story to Mac-based deployments.

Sources