Qwen3.6 35B A3B FP8
Multimodal agentic coding model with thinking preservation and 262K context
About model
Qwen3.6 35B A3B is the first open-weight release in Qwen's 3.6 series, a 35B parameter Mixture-of-Experts model that activates 3B parameters per token, served in FP8 on Together AI. It pairs a hybrid design combining Gated DeltaNet linear attention with gated attention layers, vision input, and 262K native context extensible up to 1M tokens, plus a new thinking preservation option that retains reasoning context from prior messages for iterative development. The release centers on agentic coding, reaching 73.4% on SWE-bench Verified, 80.4% on LiveCodeBench v6, and 51.5% on Terminal-Bench 2.0, with stronger frontend workflows and repository-level reasoning.
73.4%
Agentic coding across real-world software engineering tasks
262K
Extensible up to 1M tokens for repository-level reasoning
35B
Hybrid MoE architecture served in FP8 for efficient inference
- Agentic Coding: 73.4% SWE-bench Verified and 51.5% Terminal-Bench 2.0, with improved frontend workflows and repository-level reasoning
- Multimodal Understanding: Vision input with 81.7% MMMU and 86.4% MathVista (mini) for document, chart, and real-world image reasoning
- Thinking Preservation: Optional retention of reasoning context from prior messages, reducing overhead in iterative multi-turn development
- Production-Ready Infrastructure: 99.9% SLA, available on Together AI dedicated infrastructure
Model | AIME 2025 | GPQA Diamond | HLE | LiveCodeBench | MATH500 | SWE-bench verified |
|---|---|---|---|---|---|---|
Qwen3.6 35B A3B FP8 | 21.4 | 80.4 | 73.4 | Related open-source models | Competitor closed-source models | |
 Claude Opus 4.6 | 90.5% | 34.2% | 78.7% | |||
 OpenAI o3 | 83.3% | 24.9% | 99.2% | 62.3% | ||
OpenAI o1 | 76.8% | 96.4% | 48.9% | |||
GPT-4o | 49.2% | 2.7% | 32.3% | 89.3% | 31.0% |
Model card
Architecture Overview:
• 35B total parameters with 3B activated per token across 256 experts (8 routed plus 1 shared)
• Hybrid 40-layer layout interleaving Gated DeltaNet linear attention with gated attention
• 262,144-token native context, extensible up to 1,010,000 tokens
• Multi-token prediction trained for faster decoding, served in FP8 on Together AI
Training Methodology:
• Pre-training and post-training with a vision encoder for image understanding
• Built on community feedback from the Qwen3.5 series, prioritizing stability and real-world coding utility
Performance Characteristics:
• Coding and agentic: 73.4% SWE-bench Verified, 67.2% SWE-bench Multilingual, 51.5% Terminal-Bench 2.0, 80.4% LiveCodeBench v6
• Reasoning and knowledge: 86.0% GPQA, 92.7% AIME 2026, 85.2% MMLU-Pro, 21.4% Humanity's Last Exam
• Vision: 81.7% MMMU, 86.4% MathVista (mini), 85.3% RealWorldQA
Prompting
Together AI API Access:
• Access Qwen3.6 35B A3B FP8 via Together AI APIs using the endpoint Qwen/Qwen3.6-35B-A3B-FP8
• Authenticate using your Together AI API key in request headers
• Supports native function calling and JSON mode for structured outputs
• Optional thinking preservation retains reasoning context from historical messages for multi-turn agentic work
• Available on Together AI dedicated infrastructure
Applications & use cases
Agentic Software Development:
• Resolve issues and pull requests across real repositories with tool-calling agent loops
• Build frontend features with the release's improved frontend workflow handling
• Keep reasoning context across turns with thinking preservation for iterative debugging
Multimodal Document & UI Understanding:
• Parse screenshots, charts, and documents as part of coding and analysis workflows
• Ground UI-generation tasks in reference images passed through the Together API
• Combine vision input with function calling for end-to-end multimodal agents
Long-Context Repository Work:
• Load large codebases into the 262K native context for repository-level reasoning
• Extend to 1M tokens for cross-document analysis and multi-file refactors
• Run sustained multi-step agent sessions without losing earlier context
