Qwen3.8-27B

2026-08-13 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba releases Qwen3.8-27B, a dense multimodal model with 27 billion parameters under the Apache 2.0 license, which outperforms the closed Qwen3.7-Plus on several agentic tests according to the vendor's figures.

### Context Qwen's open models moved from Qwen3.6 (April) to Qwen3.8, while the Qwen3.7 generation (Max and Plus) remained proprietary and API-only. The August 3 announcement promised two open releases, Max and 27B. The 27B was released after a slight delay of a few days relative to the announced week (third-party sources). ### What the model brings - Agentic coding: SWE-bench Pro 61.7 vs 53.5 for Qwen3.6-27B and 57.6 for Qwen3.7-Plus (closed), vendor scores. - Computer use: OSWorld-Verified 84.3 vs 73.3 for Qwen3.7-Plus and 72.7 for Opus 4.6 Max. - Thinking control: thinking mode by default, `reasoning_effort`, `preserve_thinking`. - Vision: images, videos, documents and long videos. - A jump in DeepSWE 1.1 from 13.3 (Qwen3.6-27B) to 42.2 within the same dense generation. ### Architecture and specifications - Dense, 27B (28B with the vision encoder according to third-party sources; 27.8 billion parameters counted on Hugging Face); 64 layers; hidden dimension 5120; layout 16 x (3 x (Gated DeltaNet then FFN) then 1 x (Gated Attention then FFN)). - Context: 262,144 tokens natively, extensible up to 1,000,000 (YaRN). - Modalities: text, image, video; text output. - License: Apache 2.0. Weights: Hugging Face, Qwen/Qwen3.8-27B (and FP8). Knowledge cutoff: not disclosed. ### Cost and access | Item | Value | |---|---| | Qwen Cloud API, input (USD / M tokens) | 0.5 | | Qwen Cloud API, output (USD / M tokens) | 3 | | Implicit cache / explicit read | 0.1 / 0.05 | | API context | 1M tokens (991K input, 131K output) | | Self-hosting | FP8 weights published; minimum hardware not disclosed | Pricing recorded on 2026-09-29 on Qwen Cloud. ### Benchmarks | Test | Qwen3.8-27B | Qwen3.6-27B | Qwen3.7-Plus (closed) | Opus 4.6 Max (closed) | |---|---|---|---|---| | Terminal Bench 2.1 | 73.0 | 63.4 | 64.0 | 78.2 | | SWE-bench Pro | 61.7 | 53.5 | 57.6 | 53.4 | | CoWorkBench | 70.7 | 61.0 | 65.1 | 68.2 | | GPQA Diamond | 89.2 | 87.8 | 90.3 | 91.3 | | HLE | 30.8 | 24.0 | 34.7 | 40.0 | | LiveCodeBench v6 | 90.3 | 83.9 | 89.6 | 88.8 | | OSWorld-Verified | 84.3 | 63.9 | 73.3 | 72.7 | Source: Hugging Face model card for Qwen3.8-27B, scores published by the vendor. ### Versus closed models The 27B outperforms Opus 4.6 Max on SWE-bench Pro (61.7 vs 53.4), CoWorkBench (70.7 vs 68.2), LiveCodeBench v6 (90.3 vs 88.8) and OSWorld-Verified (84.3 vs 72.7). It remains behind on Terminal Bench 2.1 (73.0 vs 78.2), GPQA Diamond (89.2 vs 91.3) and especially HLE (30.8 vs 40.0). Against Qwen3.7-Plus, closed and from the same company, it wins on agentic coding and loses on HLE and GPQA. The Opus 4.6 Max comparator is a closed model older than the most recent Opus and GPT-5.x models of the period, cited in the Qwen3.8-Max card. The API is priced at 0.5 and 3 USD per million tokens, and the model can be self-hosted under Apache 2.0: it is the freest open variant of the 3.8 generation. ### Key takeaways A 27-billion-parameter model under Apache 2.0 matches or outperforms closed models of the previous generation on agentic coding and computer use. It remains behind on pure knowledge reasoning (HLE).

Tags : LLM, Open Source, Alibaba, Qwen, Apache 2.0, Multimodal, Modèle dense, Agentique