Overview
Qwen3.8-27B is the compact dense model in Alibaba's Qwen3.8 open-weight generation, published on Hugging Face in August 2026 under Apache 2.0 alongside an FP8 checkpoint. Qwen's model card introduces Qwen3.8 as 'the most capable generation in the Qwen open-model family to date' and frames the 27B as bringing those advances 'to a compact, deployment-friendly dense model: a native vision-language model that understands images and videos, with flexible thinking control'.
It is a 27-billion-parameter causal language model with a vision encoder, built on the architectural foundation of Qwen3.5. The 64-layer stack interleaves Gated DeltaNet linear-attention blocks with Gated Attention blocks, each followed by a feed-forward network, and is trained with multi-token prediction. Context length is 262,144 tokens natively and extensible to 1,000,000 with YaRN scaling; Qwen recommends allocating up to 262,144 tokens for reasoning content and 131,072 for the final response.
Thinking mode is on by default and can be disabled per request. Reasoning depth is tuned with `reasoning_effort` (xhigh by default, plus medium and low) and reasoning context from earlier turns is retained with `preserve_thinking`. Qwen recommends temperature 1.0 / top_p 0.95 / top_k 20 in thinking mode and temperature 0.7 / top_p 0.80 / top_k 20 in instruct mode.
In Qwen's own launch table the 27B leads its predecessor Qwen3.6-27B and the hosted Qwen3.7-Plus on most coding and agent benchmarks — Terminal Bench 2.1 73.0 (vs 63.4 and 64.0), SWE-bench Pro 61.7, QwenSWEBench 79.0, DeepSWE 1.1 42.2 — and posts strong multimodal agent numbers, including OSWorld-Verified 84.3 for computer use and AndroidWorld 81.9 for mobile use. Qwen evaluates most coding rows with the Claude Code harness and marks unavailable cells as '--'.
The weights ship in Hugging Face Transformers format and run on SGLang, vLLM and TokenSpeed, each with a published Qwen3.8 recipe. Qwen says a hosted Qwen3.8-27B with production features such as a 1M default context and built-in tools is coming to Qwen Cloud.
| Released | 2026-08 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Parameters | 27B dense |
| Context | 262K (extensible to 1M) |
| Max output | 131K final response · 262K reasoning |
| Architecture | Causal language model with a vision encoder, built on the architectural foundation of Qwen3.5. 64 layers in a hybrid layout — 16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN)) — with a 5120 hidden dimension, 17,408 FFN intermediate dimension, 48 linear-attention value heads and 16 QK heads in the Gated DeltaNet blocks, 24 query and 4 key-value heads in the Gated Attention blocks, and multi-token prediction trained with multiple steps. |
| Modalities | Text, Vision, Video |
| Status | Available (open weights on Hugging Face). Qwen lists a hosted Qwen Cloud version as coming soon. |
Benchmarks
Qwen3.8-27B against the peers Qwen names in its own model-card tables. Coding rows are evaluated with the Claude Code harness; MathVision, BabyVision and CharXiv (RQ) rows are the "without code interpreter" setting. Blank cells are the '--' Qwen marks as not available.
| Benchmark | Qwen3.8-27B | Qwen3.6-27B | Qwen3.7-Plus | Muse Glimmer-30B | Opus4.6 Max |
|---|---|---|---|---|---|
| Terminal Bench 2.1 (Terminus) | 73% | 63.4% | 64% | 51.7% | 78.2% |
| SWE-bench Pro | 61.7% | 53.5% | 57.6% | 51.2% | 53.4% |
| NL2Repo-Bench | 42.3% | 36.2% | 41.1% | — | 47.6% |
| DeepSWE 1.1 | 42.2% | 13.3% | 14.2% | — | — |
| QwenSWEBench | 79% | 49.3% | 59.2% | — | 63.8% |
| CoWorkBench | 70.7% | 61% | 65.1% | — | 68.2% |
| JobBench | 33.4% | 21.8% | 27.6% | — | — |
| Agents' Last Exam (Pass@1) | 20.4% | 10.6% | 13.2% | — | — |
| IFBench | 79.5% | 69.1% | 79.1% | 77% | 62.5% |
| GPQA Diamond | 89.2% | 87.8% | 90.3% | 83.5% | 91.3% |
| HLE | 30.8% | 24% | 34.7% | 22% | 40% |
| LiveCodeBench v6 | 90.3% | 83.9% | 89.6% | — | 88.8% |
| OSWorld-Verified (computer use) | 84.3% | 63.9% | 73.3% | 65.9% | 72.7% |
| WebArena-Verified (browser use) | 64.8% | 48.8% | 55.3% | — | — |
| AndroidWorld (mobile use) | 81.9% | 70.3% | 81% | — | 62% |
| RecreationBench | 47.1% | 29.8% | 30.2% | — | — |
| SWE-MM (multimodal SWE) | 38.6% | 25.7% | 30% | — | 27.1% |
| Vision2Web | 62.9% | 45% | 42.1% | — | — |
| MathVision (without CI) | 90% | 85.1% | 90.3% | — | 65.5% |
| BabyVision (without CI) | 65.7% | 28.9% | 64.7% | — | 12.6% |
| CharXiv RQ (without CI) | 83.7% | 78.4% | 85.8% | 78.8% | 66% |
| OmniDocBench 1.5 | 91.1% | 89.4% | 91.4% | 75.8% | 86.6% |
| RealWorldQA | 85.9% | 84.1% | 86.9% | — | 73.9% |
| ERQA | 65.5% | 62.5% | 69.8% | — | 40.8% |
This model's scores
- Terminal Bench 2.1 (Terminus)73%
- SWE-bench Pro61.7%
- QwenSWEBench79%
- LiveCodeBench v690.3%
- GPQA Diamond89.2%
- IFBench79.5%
- OSWorld-Verified (computer use)84.3%
- AndroidWorld (mobile use)81.9%
- OmniDocBench 1.591.1%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Terminal Bench 2.1 of 73.0 for agentic terminal coding, against 63.4 for Qwen3.6-27B and 64.0 for Qwen3.7-Plus
- QwenSWEBench 79.0 and DeepSWE 1.1 42.2 — large jumps over the 49.3 and 13.3 Qwen reports for Qwen3.6-27B
- OSWorld-Verified 84.3, WebArena-Verified 64.8 and AndroidWorld 81.9 for computer, browser and mobile agent use
- Native image and video understanding, from STEM diagrams and documents to hour-scale video
- 262K native context extensible to 1M, in a 27B dense model that fits far smaller deployments than the 2.4T MoE sibling
- Apache 2.0 weights, so commercial self-hosting carries no custom-licence conditions
Best for
- Self-hosted coding and terminal agents where a 27B dense model is the largest that fits the hardware budget
- Computer-, browser- and mobile-use agents that need vision grounded in the same model
- Document and chart understanding — OmniDocBench 1.5 91.1 and CharXiv (RQ) 83.7 without code interpreter
- Long-context work over repositories or hour-scale video within the 262K–1M window
Qwen (open-weight) — every version
The full lineage of the Qwen (open-weight) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Qwen3.8-2.4T-A95Bcurrent | 2026-08 | 262K (up to 1M) | Qwen3.8-Max License |
| Qwen3.8-27B | 2026-08 | 262K (up to 1M) | Apache-2.0 |
| Qwen3.6 | 2026-04 | — | Apache-2.0 |
| Qwen3.5 | 2026-02-16 | — | Apache-2.0 |
| Qwen3 (2507 update) | 2025-07 | — | Apache-2.0 |
| Qwen3 | 2025-04-28 | — | Apache-2.0 |
| Qwen2.5 | 2024-09 | — | Apache-2.0 |
| Qwen2 | 2024-06 | — | Apache-2.0 |
FAQ
What licence are the Qwen3.8-27B weights under?
Apache 2.0, as published on the Hugging Face model card. That is more permissive than the custom licence on the larger Qwen3.8-2.4T-A95B checkpoint.
Is Qwen3.8-27B multimodal?
Yes. It is a causal language model with a vision encoder and natively understands images and video, from STEM diagrams and documents to hour-scale videos, in addition to text.
What context length does Qwen3.8-27B support?
262,144 tokens natively, extensible up to 1,000,000 with YaRN scaling. Qwen recommends allocating up to 262,144 tokens to reasoning content and 131,072 to the final response.
Can thinking mode be turned off?
Yes. Thinking mode is on by default but can be disabled per request. Reasoning depth is set with reasoning_effort (xhigh by default, plus medium and low), and reasoning context from earlier messages is retained with preserve_thinking.
How does Qwen3.8-27B compare with Qwen3.6-27B?
On Qwen's own tables it leads across the board: Terminal Bench 2.1 73.0 vs 63.4, SWE-bench Pro 61.7 vs 53.5, QwenSWEBench 79.0 vs 49.3, DeepSWE 1.1 42.2 vs 13.3, OSWorld-Verified 84.3 vs 63.9 and IFBench 79.5 vs 69.1.
How do I run Qwen3.8-27B?
The weights ship in Hugging Face Transformers format and are compatible with SGLang, vLLM and TokenSpeed, each of which publishes a Qwen3.8 recipe. An FP8 checkpoint is also published for lower-memory serving.