TrueCourse/Qwen3.6-35B-AWQ
Qwen3.6-35B-AWQ is a 35.1 billion parameter causal language model developed by Qwen, featuring a Mixture-of-Experts (MoE) architecture with 3 billion activated parameters and a native context length of 262,144 tokens, extensible up to 1,010,000 tokens. This model is optimized for agentic coding, excelling in frontend workflows and repository-level reasoning, and includes a vision encoder for multimodal capabilities. It introduces a 'Thinking Preservation' feature to retain reasoning context from historical messages, enhancing iterative development and decision consistency.
Loading preview...
Qwen3.6-35B-AWQ: An Agentic Coding and Multimodal LLM
Qwen3.6-35B-AWQ is a 35.1 billion parameter causal language model from the Qwen3.6 series, featuring a Mixture-of-Experts (MoE) architecture with 3 billion activated parameters. It boasts a native context length of 262,144 tokens, which can be extended up to 1,010,000 tokens using YaRN scaling techniques. This model is designed for enhanced stability and real-world utility, particularly in developer workflows.
Key Capabilities and Differentiators
- Agentic Coding: Significantly upgraded capabilities in handling frontend workflows and repository-level reasoning, demonstrated by strong performance on benchmarks like SWE-bench Verified (73.4) and Terminal-Bench 2.0 (51.5).
- Thinking Preservation: A unique feature allowing the model to retain reasoning context from historical messages, which streamlines iterative development, reduces overhead, and improves decision consistency in agent scenarios.
- Multimodal Input: Supports image and video inputs, functioning as a Vision Language Model with a dedicated vision encoder. It shows competitive performance across various vision benchmarks, including MMMU (81.7) and RealWorldQA (85.3).
- Ultra-Long Context: Natively supports a 262,144-token context window, with extensibility up to 1,010,000 tokens, making it suitable for complex, long-horizon tasks.
- Optimized Inference: Compatible with high-throughput inference frameworks like SGLang and vLLM, with specific configurations for tool use and Multi-Token Prediction (MTP).
Good for
- Code Generation and Debugging: Developers requiring advanced agentic coding capabilities, including frontend development and repository-level code understanding.
- Complex Agent Workflows: Scenarios where maintaining reasoning context across multiple turns is crucial for consistent decision-making and reduced token consumption.
- Multimodal Applications: Tasks involving both text and visual (image or video) inputs, such as visual question answering, document understanding, and video analysis.
- Long-Context Processing: Applications that demand processing and generating responses for extremely long texts, leveraging its extended context window.