News · High impactBack to News

Alibaba Cloud Open-Sources Qwen3.8-27B: Hybrid Attention and Single-GPU Long-Horizon Agents

Alibaba Cloud's Qwen team has open-sourced Qwen3.8-27B, a dense 27B vision-language model. Built on a hybrid architecture combining 48 linear attention layers with 16 gated attention layers, it features a native MTP draft head and controllable thinking mode, delivering native 262K context and single-GPU deployment while advancing SWE-bench Pro and LiveCodeBench benchmarks.

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

  1. Ingest Multimodal Inputs and Environmental Feedback

    Parse structured text, high-resolution charts, and code execution states returned by tools.

  2. Hierarchical Autonomous Reasoning and Planning

    Generate reflective reasoning traces according to configured effort levels, decomposing tasks and validating preconditions.

  3. 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).

BenchmarkDomainQwen3.8-27BQwen3.6-27BQwen3.7-PlusMuse Glimmer-30BOpus4.6 Max
SWE-bench ProAgentic Coding (SWE)61.753.5--54.153.4
QwenSWEBenchSoftware Engineering79.049.359.2--63.8
LiveCodeBench v6Competitive Coding90.383.989.6--88.8
Terminal Bench 2.1Terminal Coding73.063.464.051.778.2
CoWorkBenchLong-Horizon Work70.761.065.1--68.2
IFBenchInstruction Following79.569.179.177.062.5
GPQA DiamondScientific Reasoning89.287.890.383.591.3

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.

Core Conclusion

Next step

Keep tracking Alibaba

Continue along the same topic.

Open entity record