inclusionAI: Ring-2.6-1T
🧠 AI Modelinclusionai
1T-parameter thinking model with 63B active params, optimized for coding agents and tool use.
Ring-2.6-1T is a large-scale MoE-style thinking model from INCLUSIONAI, leveraging 1 trillion total parameters while activating only 63B per forward pass. This architecture reduces computational cost while maintaining high reasoning capability. It supports a context length of 262,144 tokens, enabling long-document and multi-step agent tasks. The model is optimized for coding agents, tool calling, and structured outputs through features like frequency/penalty parameters, reasoning tokens, and response format control. Pricing is aggressive: $0.07 per million input tokens and $0.62 per million output tokens. Benchmark performance includes AIME 2025 (65.3%), GPQA Diamond (60.2%), MATH-500 (76.4%), MMLU (84.6%), and SimpleQA (72.4%). These scores place it competitively against GPT-4 and Claude 3.5, despite lower active parameters and cost.
💡Highlights
- ├─1T total params, 63B active per inference
- ├─262k context window for long tasks
- └─Top scores on AIME 2025 & GPQA Diamond
🎯For
- ├─AI researchers
- ├─Developers
- └─Agent builders