Qwen 3.8 27B: Next-Gen Open-Source Language Model
Delivers significant improvements in coding, professional tasks, and research domains.
1Qwen 3.8 27B Announced
Qwen introduced its new open-source language model family member, Qwen 3.8 27B. Following the widespread adoption of Qwen 3.5 and 3.6 series, this model offers substantial gains in coding, professional work, research, and long-horizon tasks. With 27 billion parameters, it's designed as a native vision-language model capable of understanding images and videos while maintaining compact deployment size.
2FP8 Quantization Technology
Qwen 3.8 27B employs fine-grained FP8 quantization with 128 block size, maintaining original performance while reducing model size. Compatible with Hugging Face Transformers, vLLM, SGLang, and TokenSpeed, the model will also be available through Qwen Cloud API service for managed, scalable inference, supporting up to 1 million tokens context length by default with production-ready features.
3Architecture and Technical Specs
The model features 64 layers with 5120 hidden dimensions. It combines Gated DeltaNet (48 V heads and 16 QK heads) with Gated Attention mechanism (24 Q heads and 4 KV heads). Native context length of 262,144 tokens can be extended to 1 million. FFN intermediate dimension is set at 17,408, while padded token embedding uses 248,320 dimensions with multi-token prediction training.
4Coding Performance
The model scored 73.0 on Terminal Bench 2.1, achieved 61.7% on SWE-bench Pro, and 42.3% on NL2Repo-Bench. It demonstrated impressive 79.0% performance on QwenSWEBench, showing clear superiority over previous generations in software engineering tasks. DeepSWE 1.1 results at 42.2% further confirm its coding capabilities.
5Professional Task Handling
Scoring 70.7% on CoWorkBench and 33.4% on JobBench, the model excels in long-horizon tasks across computer science, finance, law, and medical domains. It achieved 20.4% Pass@1 and 42.9 score on Agents' Last Exam for frontier tasks, while leading general instruction following with 79.5% on IFBench.
6Scientific Reasoning
The model scored 89.2% on GPQA Diamond and 30.8% on HLE for multidisciplinary reasoning. It dominates competitive coding with 90.3% on LiveCodeBench v6. For mathematical problem solving, it achieves 90.0% without CI and 94.6% with CI on MathVision, demonstrating strong analytical capabilities.
7Multimodal Capabilities
Performance includes 84.3% on OSWorld-Verified, 64.8% on WebArena-Verified, and 81.9% on AndroidWorld. It scores 47.1% on RecreationBench and 57.4 Pass@3 average on ClawEval-MM for multimodal tool use. The model shows particular strength in visual software engineering with 38.6% on SWE-MM and 62.9% on Vision2Web.
8Visual Understanding
The model achieves 65.7% without CI and 85.6% with CI on BabyVision. For scientific chart analysis, it scores 83.7% without CI and 90.2% with CI on CharXiv (RQ). It demonstrates strong real-world perception with 91.1% on OmniDocBench 1.5 and 85.9% on RealWorldQA, plus 65.5% on ERQA for embodied intelligence tasks.
9Training and Benchmarking
All evaluations used Claude Code harness at temp=1.0, top_p=0.95 with 256K context window. Problematic tasks were corrected, and baseline models were re-evaluated on refined benchmarks. Reward hacking prevention measures included disabling specific repository access commands like pip download, pip install, and git clone for NL2Repo-Bench.
10Production Features
The managed version through Qwen Cloud will offer production features including default 1M token context length and official built-in tools. Designed to reliably complete complex, multi-step tasks, the model features flexible thinking control and native vision-language capabilities for image and video understanding, making it particularly suitable for enterprise applications.
11Performance Comparison
Qwen 3.8-27B shows clear advantages over Qwen 3.6-27B and Qwen 3.7-Plus across all benchmarks. It also leads against competitors like Muse, Glimmer-30B, and Opus4.6 Max in multiple domains. The performance gap is particularly notable in coding and software engineering tasks, where it establishes new standards for open-weight models.
12Industry Impact
The release of Qwen 3.8 27B could mark a significant milestone in the open-source language model ecosystem. Its potential to automate professional workflows may revolutionize productivity. By accelerating R&D processes, it could support scientific discoveries. The extended context capability might set new standards for complex document analysis, while its multimodal understanding opens new application possibilities across industries.