Evaluating Qwen3.8 27B Coding Performance for Developers

On Qwen's own benchmarks, Qwen3.8-27B delivers competitive results for software development. The model scores 61.7 on SWE-bench Pro and 90.3 on LiveCodeBench v6. It advances beyond Qwen3.6-27B across agentic tasks, but trails Opus4.6 Max on Terminal Bench 2.1 with 73.0 compared to 78.2. Because these scores are vendor-reported, teams evaluating Qwen3.8 27B coding workflows must test against internal codebases.

The release provides an open-weight dense model under the Apache-2.0 license. It combines 27 billion parameters with a native vision encoder. The native context window spans 262,144 tokens. RoPE scaling extends that capacity to 1,000,000 tokens for long-context tasks.

Developers have discussed the model on r/LocalLLaMA. However, the model card omits hardware requirements and minimum VRAM figures. Engineering teams must measure memory footprint and throughput on their own target machines.

Is Qwen3.8-27B Good for Coding?

Qwen3.8-27B scores above Qwen3.6-27B on every coding row of the card's table, while trailing Opus4.6 Max on Terminal Bench 2.1 and NL2Repo-Bench. On Qwen's own benchmarks, the model leads the evaluated comparison set on SWE-bench Pro with 61.7 and LiveCodeBench v6 with 90.3. It also shows substantial gains over Qwen3.6-27B on agentic coding, scoring 42.2 on DeepSWE 1.1 compared to 13.3.

The card says SWE-bench Pro, NL2Repo-Bench, DeepSWE 1.1 and QwenSWEBench were evaluated with the Claude Code harness. Every figure is vendor-reported.

• Terminal Bench 2.1 (Terminus): Qwen3.8-27B scores 73.0, trailing Opus4.6 Max at 78.2 while leading Qwen3.7-Plus at 64.0, Qwen3.6-27B at 63.4, and Muse Glimmer-30B at 51.7.
• SWE-bench Pro: Qwen3.8-27B achieves 61.7, leading Qwen3.7-Plus at 57.6, Qwen3.6-27B at 53.5, Opus4.6 Max at 53.4, and Muse Glimmer-30B at 51.2.
• NL2Repo-Bench: Qwen3.8-27B scores 42.3 on repository code generation, behind Opus4.6 Max at 47.6, but ahead of Qwen3.7-Plus at 41.1 and Qwen3.6-27B at 36.2.
• DeepSWE 1.1: Qwen3.8-27B reaches 42.2 on agentic coding, compared to 14.2 for Qwen3.7-Plus and 13.3 for Qwen3.6-27B.
• QwenSWEBench: Qwen3.8-27B scores 79.0 on the Qwen team's in-house benchmark, outperforming Opus4.6 Max at 63.8, Qwen3.7-Plus at 59.2, and Qwen3.6-27B at 49.3.
• LiveCodeBench v6: Qwen3.8-27B scores 90.3 in competitive coding, surpassing Qwen3.7-Plus at 89.6, Opus4.6 Max at 88.8, and Qwen3.6-27B at 83.9.

Qwen3.8 27B Coding Benchmark Methodology and Evaluation Limits

The coding benchmarks published by the Qwen team follow specific execution protocols. The team evaluated SWE-bench Pro, NL2Repo-Bench, DeepSWE 1.1, and QwenSWEBench using the Claude Code harness. For both SWE-bench Pro and DeepSWE 1.1, the evaluation framework applied temperature 1.0, top_p 0.95, and a 256K context window.

All baseline models were re-evaluated on this refined benchmark suite. The officially reported score for Opus4.6 Max serves as the baseline on SWE-bench Pro. Because the scores are vendor-reported, test the model on your own repositories before relying on them.

Recommended Qwen3.8 27B Coding Settings and Thinking Controls

Thinking mode is on by default and can be disabled per request. Developers can set reasoning_effort to xhigh (the default, for complex tasks demanding thorough analysis), medium (balancing accuracy and speed), or low (efficient reasoning optimizing for speed and cost).

The card warns that in multi-turn agentic tasks, lower reasoning effort does not always reduce overall task completion time. Faster per-turn responses can lead to insufficient analysis, more failures, and repeated retries, which may increase total latency and token consumption.

The card lists recommended sampling parameters for each mode.

• Thinking mode sampling: Set temperature to 1.0, top_p to 0.95, top_k to 20, min_p to 0.0, presence_penalty to 0.0, and repetition_penalty to 1.0.
• Instruct mode sampling: When disabling thinking mode, set temperature to 0.7, top_p to 0.80, top_k to 20, min_p to 0.0, presence_penalty to 1.5, and repetition_penalty to 1.0.
• Recommended output lengths: Within the 1M context, reasoning content up to 262,144 tokens and a final response up to 131,072 tokens.
• State preservation: preserve_thinking is on by default.

Downstream Tooling Support and Dedicated Serving Frameworks

The card lists downstream compatibility among its highlights, with broader support for popular harnesses and development tools. The checkpoint works with Hugging Face Transformers, vLLM, SGLang, and TokenSpeed. For production workflows, the Qwen team explicitly recommends dedicated serving engines like SGLang, vLLM, or TokenSpeed.

Inference efficiency varies significantly across frameworks. Teams can access local instances using the OpenAI Python SDK and Chat Completions API style, targeting model identifier 'Qwen/Qwen3.8-27B'.

Vision Model Capabilities for Interface and Document Automation

Qwen3.8-27B includes a native vision encoder, and the card describes image and video understanding, from STEM diagrams and documents to hour-scale videos.

The model card reports competitive results on computer and browser benchmarks. These vision scores are higher than Qwen3.6-27B's on every row listed.

• OSWorld-Verified (computer use): Qwen3.8-27B scores 84.3, outperforming Opus4.6 Max at 72.7, Qwen3.7-Plus at 73.3, and Qwen3.6-27B at 63.9.
• WebArena-Verified (browser use): Qwen3.8-27B achieves 64.8, compared to 55.3 for Qwen3.7-Plus and 48.8 for Qwen3.6-27B.
• AndroidWorld (mobile device tasks): Qwen3.8-27B reaches 81.9, ahead of Qwen3.7-Plus at 81.0, Qwen3.6-27B at 70.3, and Opus4.6 Max at 62.0.
• OmniDocBench 1.5 (document intelligence): Qwen3.8-27B scores 91.1, ahead of Qwen3.6-27B at 89.4 and Opus4.6 Max at 86.6 but behind Qwen3.7-Plus at 91.4.
• RealWorldQA: Qwen3.8-27B scores 85.9, ahead of Qwen3.6-27B at 84.1 and Opus4.6 Max at 73.9 but behind Qwen3.7-Plus at 86.9.

Open Weight Licensing and Deployment Hardware Considerations

The open weights for Qwen3.8-27B are published under the Apache-2.0 license, and the weights are free to download. The Qwen team's Hugging Face repository was created on August 5, 2026. The architecture contains 27 billion parameters across 64 layers with a hidden dimension of 5120. Its hidden layout repeats 16 times (three Gated DeltaNet layers, then one Gated Attention layer), and it was trained with multi-token prediction (MTP).

The model card publishes no hardware specifications, minimum VRAM figures, or recommended GPU counts. Engineering teams must measure memory use on their own target hardware. The Qwen team has announced that an official hosted version with a 1,000,000 token default context window is coming soon to Qwen Cloud.

Hosted pricing remains unpublished as of October 2026.

Engineering Steps to Test Qwen3.8 27B Coding on Internal Repositories

Software engineering teams should evaluate Qwen3.8-27B by running offline benchmarks against historical pull requests and real defect reports. A Reddit discussion titled 'Engineer / developer observations of Gemma4-31B, Qwen3.8-27B, and 6.1-Sol for software engineering work' is one sign that developers are comparing the model.

To test real utility, compile an evaluation set of historical tickets with test suites. Run Qwen3.8-27B with thinking mode on and the default xhigh reasoning effort. Download the weights from Hugging Face, serve them with one of the engines named above, and score the results on your own internal pull requests.