Hybrid Attention Architecture and Native Speculative Decoding
Alibaba Cloud's Qwen team has officially open-sourced Qwen3.8-27B, a dense vision-language foundation model. Positioned as the developer-friendly, mid-sized anchor of the Qwen3.8 generation under the Apache 2.0 license, the model provides an efficient, low-memory open base for coding, research, and complex agent workflows.
Core Architecture and Performance Specifications
-
27B
Parameter dense vision-language model, evolved on the Qwen3.5 architecture.
-
48/64
Layers utilize linear attention, drastically cutting KV cache footprint and long-context latency.
-
262K
Native context window tokens, extensible up to 1M (100万) tokens via RoPE scaling.
Qwen3.8-27B adopts a hybrid design combining 48 linear attention layers with 16 gated attention layers, preserving deep reasoning capacity while dramatically reducing long-context memory overhead. Additionally, the model integrates a native Multi-Token Prediction (MTP) draft head, enabling out-of-the-box speculative decoding speedups in inference engines such as vLLM and SGLang without secondary draft models.
Controllable Thinking Mode and Long-Horizon Continuity
For reasoning control, Qwen3.8-27B activates built-in thinking mode by default, allowing the model to generate multi-step reasoning traces prior to its final response. Developers can tune reasoning depth via API parameters across xhigh, medium, and low levels, balancing inference accuracy against latency.
Long-Horizon Agent Reasoning and Execution Pipeline
-
Ingest Multimodal Inputs and Environmental Feedback
Parse structured text, high-resolution charts, and code execution states returned by tools.
-
Hierarchical Autonomous Reasoning and Planning
Generate reflective reasoning traces according to configured effort levels, decomposing tasks and validating preconditions.
-
Preserve Reasoning Context and Execute Actions
Maintain thinking continuity across multi-turn interactions, accurately invoking external APIs and terminal environments.
The model enables cross-turn reasoning preservation by default, ensuring agents maintain previous reasoning context across multi-step environmental interactions and tool invocations. Coupled with a native vision encoder supporting charts, technical documents, and hour-scale video, the model reliably executes complex engineering troubleshooting and autonomous research workflows.
Benchmark Advancements and Cross-Model Comparisons
In official benchmark evaluations, Qwen3.8-27B demonstrates engineering capabilities that surpass comparable models and frontier flagships. On SWE-bench Pro, it achieved 61.7, substantially outperforming Anthropic's flagship Opus4.6 Max (53.4) and predecessor Qwen3.6-27B (53.5); on the rigorous software engineering benchmark QwenSWEBench, it reached 79.0, leading Opus4.6 Max (63.8).
| Benchmark | Domain | Qwen3.8-27B | Qwen3.6-27B | Qwen3.7-Plus | Muse Glimmer-30B | Opus4.6 Max |
|---|---|---|---|---|---|---|
| SWE-bench Pro | Agentic Coding (SWE) | 61.7 | 53.5 | -- | 54.1 | 53.4 |
| QwenSWEBench | Software Engineering | 79.0 | 49.3 | 59.2 | -- | 63.8 |
| LiveCodeBench v6 | Competitive Coding | 90.3 | 83.9 | 89.6 | -- | 88.8 |
| Terminal Bench 2.1 | Terminal Coding | 73.0 | 63.4 | 64.0 | 51.7 | 78.2 |
| CoWorkBench | Long-Horizon Work | 70.7 | 61.0 | 65.1 | -- | 68.2 |
| IFBench | Instruction Following | 79.5 | 69.1 | 79.1 | 77.0 | 62.5 |
| GPQA Diamond | Scientific Reasoning | 89.2 | 87.8 | 90.3 | 83.5 | 91.3 |
Qwen3.8-27B consistently outperforms Opus4.6 Max across SWE-bench Pro and QwenSWEBench, setting a new precedent for lightweight models against frontier flagships. Furthermore, the model requires approximately 17GB of VRAM under 4-bit quantization, achieving Day 0 support across vLLM, SGLang, llama.cpp, LM Studio, Ollama, Unsloth, and AMD ROCm platforms.