Moonshot AI Releases Kimi K3.1 Pro With 2.8T Parameters and 1M Context

Kimi K3.1 Pro is the first model in Moonshot's K3.1 series. It is a post-trained version of Kimi K3 with much larger post-training compute. Moonshot ran joint reinforcement learning across hundreds of downstream domains and optimized its reinforcement-learning framework, long-context inference, sandbox infrastructure and rollout scheduling. The base K3 architecture uses Kimi Delta Attention and Attention Residuals, and it activates 16 of 896 experts through a Stable LatentMoE framework. The API accepts text, images and video, and Moonshot also shipped Kimi K3.1 Pro Highspeed, which the company says runs two to three times faster. MarkTechPost described the release as a 3T-class open-weight model, though it noted that the weights are not yet available.

The benchmark figures cited by MarkTechPost were self-reported by Moonshot at maximum reasoning effort, mostly with the Claude Code harness. On Terminal-Bench 4.0, K3.1 Pro scored 65.7%, 0.7 points behind Claude Opus 5.5 at 66.4% in Kimi's chart. It led SWE-Marathon at 51.7, ahead of Opus 5.5 at 46.3, and scored 61.2 on the internal Kimi Code Bench 3.0 Long-horizon, up from 55.95 for K3. The gap was larger on Terminal-Bench Science 0.1, where K3.1 Pro scored 38.1 versus 68.1 for GPT-6 Astra and 63.3 for Opus 5.5 in Kimi's chart.

On BenchCAD with tools, K3.1 Pro scored 92.7, behind Opus 5.5 at 96.2 and GPT-6 Astra at 95.9; without tools it scored 72.8, close to Opus 5.5 at 73.0. Moonshot reported 64.4 on OfficeQA Pro v2 and 58.6 on Frontier Finance, and said internal finance tests improved by 7.2 points in modeling and 11.6 points in research over K3. On the internal Kimi Hallu Bench Hard, claim-level hallucinations fell from 23.4% for K3 to 8.3% for K3.1 Pro; GPT-6 Astra posted 5.5% on the same test.

For the open-weight release, Moonshot removed cyber-related data from mid-training and reinforcement learning, according to the report. On its internal xdev-bench-hard test, an unfiltered checkpoint solved 90.5% at pass@3, while the shipped K3.1 Pro solved 81.0%. Moonshot said the model remains behind frontier proprietary models on offensive cyber tasks.

Kimi K3.1 Pro is live on Kimi.ai, Kimi Work, Kimi Code and the Kimi API. The API IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed, and the endpoint is OpenAI-compatible. Pricing is $2.00 per million input tokens and $10.00 per million output tokens, with cache hits at $0.20; Highspeed costs $4.00 input and $20.00 output. K3 was priced at $3.00 input and $15.00 output. According to the quickstart, thinking mode is always on, reasoning effort can be low, high or max with max as the default, temperature is fixed at 1.0 and top-p at 0.95, and output defaults to 131,072 tokens and can go up to 1,048,576. Access requires a minimum $1 top-up, and vision input does not accept public image URLs; developers must use base64 or uploaded file IDs.

The report noted several limitations. The weights are not public, so self-hosting is not possible today. The 2.8 trillion-parameter size points to multi-node, supernode-class serving rather than a single GPU. Science-heavy terminal work trails GPT-6 Astra and Opus 5.5 by more than 25 points, and several claimed wins come from internal benchmarks. Moonshot's K3 documentation warned that dropping reasoning history in a harness destabilizes output, and the K3.1 Pro docs say to return the full assistant message.

Editor's Summary

Moonshot AI has released Kimi K3.1 Pro, a 2.8T-parameter MoE model with 104B active parameters and a 1M-token context window. The API is available at lower prices than K3, while open weights are expected in November 2026. Moonshot's benchmark claims are self-reported and show strong coding results but a large gap on scientific terminal tasks.