darkc0de/Qwen3.8-27B-heretic
The Qwen3.8-27B model, developed by Qwen, is a 27 billion parameter causal language model with a native vision encoder, offering a 262,144 token context length, extensible to 1,000,000 tokens. It excels in agentic tasks, coding, professional work, and multimodal understanding, supporting image and video inputs. This model is designed for reliable completion of complex, multi-step tasks with flexible thinking control.
Loading preview...
Qwen3.8-27B: Advanced Multimodal Agentic LLM
Qwen3.8-27B is the latest 27 billion parameter model in the Qwen open-model family, building upon the Qwen3.5 architecture. It is a native vision-language model capable of understanding both images and videos, designed for complex, multi-step task completion with enhanced reliability.
Key Capabilities
- Comprehensive Improvements: Significant gains across coding, professional work, research, and long-horizon agentic tasks.
- Agent Execution: Features stronger autonomous planning and improved handling of environment feedback, leading to more reliable end-to-end task completion.
- Flexible Thinking Control: Operates in a 'thinking mode' by default, generating internal reasoning steps. This can be disabled, and reasoning depth (
reasoning_effort) can be tuned (xhigh, medium, low). Historical reasoning context is retained viapreserve_thinking. - Vision-Language Understanding: Natively supports image and video understanding, from STEM diagrams and documents to hour-long videos.
- Extended Context: Offers a native context length of 262,144 tokens, extensible up to 1,000,000 tokens using YaRN scaling techniques.
Performance Highlights
Qwen3.8-27B demonstrates strong performance across various benchmarks, often outperforming previous Qwen versions and comparable models:
- Coding: Achieves 61.7 on SWE-bench Pro and 79.0 on QwenSWEBench for agentic coding, and 90.3 on LiveCodeBench v6 for competitive coding.
- Agentic Tasks: Scores 70.7 on CoWorkBench for long-horizon office work and 33.4 on JobBench for professional tasks.
- Multimodal Agentic Intelligence: Leads with 84.3 on OSWorld-Verified for computer use and 64.8 on WebArena-Verified for browser use.
- General Multimodal Intelligence: Shows high scores in visual math problem solving (94.6 with CI on MathVision) and general visual reasoning (85.6 with CI on BabyVision).
Good for
- Developers building agents: Its strong agent execution, flexible thinking control, and robust handling of environment feedback make it suitable for developing sophisticated AI agents.
- Applications requiring multimodal understanding: Ideal for tasks involving analysis of images, videos, and documents, such as visual question answering, scientific chart analysis, and document intelligence.
- Complex coding and software engineering tasks: Excels in agentic coding, repo-level code generation, and software engineering benchmarks.
- Long-context applications: With a native 262K context and 1M extensibility, it's well-suited for processing and reasoning over very long texts and multi-turn conversations.