Ant Group breaks the performance barrier with Ring-1T-2.5, the world’s first hybrid linear architecture “thinking model” boasting a 1T-scale parametric design. This breakthrough delivers unprecedented efficiency without compromising intelligence, running faster than 32B models while using more than 10x less memory. Rigorous testing has demonstrated elite-level reasoning, and Ring-1T-2.5 achieved gold-level results in both the IMO 2025 (35/42) and CMO 2025 (105/126) benchmarks, even exceeding the Chinese national team standards. Designed for complex problem-solving and agent workflows, this model is comparable to leading closed-source alternatives such as Gemini-3.0-Pro Thinking and GPT-5.2 Thinking, and currently maintains open-source state-of-the-art performance across inference, math, code, and agents.
Hybrid linear architecture allows 1T parameter scaling of Ring-1T-2.5
A unique combination of attention mechanisms allows Ring-1T-2.5 to process information with incredible speed and efficiency. Ant Group’s new model accomplishes this through a “1:7 blend of Multi-Head Linear Attention (MLA) and Lightning Linear” combined with partial RoPE and QK normalization. This is a rethinking of how models focus on relevant data. This innovative architecture supports more than three times the throughput of existing systems when processing context lengths greater than 32K, significantly reducing the computational cost of complex inference tasks. Ring-1T-2.5’s design prioritizes practical applications and seamlessly integrates with both Claude Code and OpenClaw agent frameworks.
The capabilities of this model have been demonstrated by successfully building a functional TinyOS internally, powered by the ASystem RL engine, and extended to autonomous task completion. According to Ant Group, Ring-1T-2.5 boasts a 1T-scale parameter design, but operates with only 63B of active parameters during inference, significantly reducing memory usage compared to traditional 32B models.
Dense Reward RLVR Training Rival Gemini and GPT-5.2 Thoughts
Ant Group’s Ring-1T-2.5 achieves performance comparable to leading proprietary models through a proprietary training methodology, dense reward reinforcement learning (RLVR) from visual rewards. This approach allowed the 1T scale parametric model to be comparable to Gemini-3.0-Pro thinking and GPT-5.2 thinking when tackling complex problem-solving tasks, despite utilizing 63B active parameters during inference. The success of this model is not just about scale. How you leverage this scale is key, and Ant Group claims it runs faster and uses more than 10x less memory than traditional 32B models.
Rigorous testing has proven that Ring-1T-2.5 can meet academically demanding challenges, achieving a score of 35/42 on the gold medal standard IMO 2025 benchmark and 105/126 on CMO 2025, which Ant Group says “even exceeded the standards of the Chinese national team.” This suggests broad applicability beyond narrowly defined tasks.
With 1T scale parameter design and 63B active parameters during inference, Ring-1T-2.5 runs faster and uses more than 10 times less memory than the traditional 32B model.
ant group
