Qwen3 235B A22B Thinking 2507 FP8
235B-parameter MoE thinking model, 256K context, 22B activated experts, state-of-the-art reasoning performance among open-source models.
About model
Qwen3 235B A22B Thinking 2507 is a causal language model that excels in complex reasoning tasks, including logical reasoning, mathematics, and science, achieving state-of-the-art results. It features enhanced 256K long-context understanding and improved performance on academic benchmarks. Suitable for users requiring advanced thinking capabilities.
Model | FrontierMath Tier 4 | GPQA Diamond | HLE | SciCode | GDPval-AA | Terminal-Bench 2.1 | Agent Arena | FrontierCode | DeepSWE |
|---|---|---|---|---|---|---|---|---|---|
Qwen3 235B A22B Thinking 2507 FP8 | 80.1% | Related open-source models | Competitor closed-source models | ||||||
 Claude Fable 5 | 87.8% | 92.6% | 53% | 60% | 62% | 85% | 53.5% | 70% | |
Claude Opus 5 | 73.2% | 93.2% | 53% | 56% | 68% | 89% | +16.4pp | 53.4% | 74% |
 GPT-5.6 Sol | 82.9% | 94.1% | 47% | 56% | 61% | 88% | 47.5% | 73% | |
 Grok 4.5 | 24.4% | 93.1% | 40% | 54% | 51% | 82% | +4.2pp | 42.4% | 54% |
GPT-5.6 Luna | 61.0% | 91.1% | 37% | 53% | 54% | 81% | 39.8% | 67% |
Model card
Architecture Overview:
• Mixture-of-Experts transformer with 235B total parameters and 22B activated
• 94 layers with grouped query attention (64 for Q and 4 for KV)
• 128 experts with 8 activated experts per token for efficient inference
• Native 262,144 token context window for extensive document processing
Training Methodology:
• Advanced pretraining & post-training pipeline with thinking capability enhancement
• Specialized training for logical reasoning, mathematics, science, and coding tasks
• Constitutional AI training for alignment with human preferences
• Optimized for complex multi-step reasoning with increased thinking length
Performance Characteristics:
• State-of-the-art results among open-source thinking models on academic benchmarks
• Exceptional performance on AIME25, HMMT25, and LiveCodeBench evaluations
• Enhanced 256K long-context understanding capabilities
• Optimized inference with MoE architecture for computational efficiency
Applications & use cases
Advanced Reasoning Applications:
• Complex mathematical problem solving & academic research
• Scientific analysis requiring multi-step logical reasoning
• Advanced coding challenges & algorithm development
• Academic benchmarking & competitive programming
Enterprise & Professional Use:
• High-complexity business analysis & strategic planning
• Technical documentation with detailed reasoning chains
• Expert-level consultation systems requiring deep thinking
• Research & development applications in specialized domains
Educational & Research:
• Graduate-level tutoring in STEM subjects
• Research paper analysis & academic writing assistance
• Complex problem decomposition for educational purposes
• Multilingual academic support across 119+ languages
Developer Integration:
• Tool calling capabilities with Qwen-Agent framework
• OpenAI-compatible API endpoints via SGLang or vLLM
• Integration with local deployment tools (Ollama, LMStudio, llama.cpp)
• Custom reasoning workflows for specialized applications
- TypeChatReasoning
- Main use casesChatReasoningFunction Calling
- FeaturesFunction CallingJSON Mode
- DeploymentDedicated
- Parameters235B
- Activated parameters22B
- Context length256K
- Input price
$0.65 / 1M tokens
- Output price
$3.00 / 1M tokens
- Input modalitiesText
- Output modalitiesText
- ReleasedJuly 24, 2025
- Quantization levelFP8
- External link
- CategoryChat
