Moonshot AI launched Kimi K3.1 Pro and K3.1 Pro Highspeed with one-million-token contexts; full model weights are planned by November 4.

Moonshot AI introduced Kimi K3.1 Pro as the first model in its K3.1 series and described it as its most capable model to date, in the three-trillion-parameter class. Both versions support one-million-token contexts, while the company says Highspeed offers equivalent intelligence with approximately two to three times the standard version’s output speed.

The reported applications include complete software-project development, game creation, 3D modeling, financial research, and extended agent tasks. Company-reported results include 65.7% on a coding test, versus 58.2% for GPT-6 Astra, and 92.7% on a 3D-modeling test; the supplied messages do not name the benchmarks or give testing conditions.

For financial workflows, the report describes processing financial data, cross-checking documents, and conducting investment analysis. It reports approximately 4.2 times the previous generation’s performance on a business-operations test and a decline in per-statement hallucination rate from 23.4% to 8.3% on a dedicated test, rather than a universal error rate across all uses.

Both versions were reported available through Kimi.ai, Kimi Work, Kimi Code, and an application programming interface (API). The detailed release report says full model weights are planned by November 4; a separate channel’s “open-weight” description does not establish that those full weights were already available on October 23.