China's Kimi K3.1 Pro: A 2.8-Trillion-Parameter AI That Beats GPT-6 on Coding Tests

Chinese AI lab Moonshot AI has unveiled its new flagship model, Kimi K3.1 Pro, a system built on a 2.8-trillion-parameter architecture that its makers say has surpassed GPT-6 on coding capability benchmarks. The model is being released in a phased public rollout that continues through November 4.
The launch shows how fast China's AI capability is compounding. Chipmaker Cerebras has also published a technical write-up of how it serves a 3-trillion-parameter Kimi model spread across 59 of its wafers, using 4-bit weight quantization and an 8-bit memory cache to reach speeds of 1,000 to more than 2,000 tokens per second. In plain terms, that means enormous models can now answer at speeds that feel instant.
The open-model side is catching up too: a project that orchestrates five smaller Qwen models together has matched a frontier lab's score on a hard coding test, though experts note the comparison is not quite apples-to-apples, since the five-model setup gets an extra checking loop. Even with that caveat, the direction is clear — powerful coding AI is getting cheaper, faster and less locked behind one company.
Source: read more
Comments
Post a Comment