Moonshot launches Kimi K3.1 Pro: 2.8T-total / 104B-active open-weight MoE with 1M context, 65.7% on Terminal-Bench 4.0, API at $2/$10, weights by Nov 4

Background and the problem it targets

The gap between open-weight models and top proprietary models on long-horizon coding and agent work is still the main worry when teams pick a model: closed models are stronger but expensive and send data off-site, while open models are cheap and self-hostable but tend to break down on long tasks. Moonshot's Kimi K3 (2.8T parameters, September) already pushed open models to the 3T class. K3.1 Pro is the first upgrade in that line, aiming to bring coding, 3D and finance agent skills close to frontier closed models on the same base while cutting API prices.

Architecture and how it works

According to the official blog, K3.1 Pro is a post-trained version of K3 on the same MoE base: 2.8T total parameters, 104B active per token, a 1M-token context and multimodal input. The gains come mainly from much larger post-training compute: Moonshot optimized the RL framework, long-context inference, training efficiency, sandboxes, rollout scheduling and an environment quality-check platform, then ran joint RL across hundreds of downstream domains. For the open-weight release, cyber-related data was removed from both mid-training and RL to limit offensive exploit capability.

Benchmarks (vendor-reported)

All figures are vendor-reported, with reasoning effort set to max: 65.7% on Terminal-Bench 4.0 versus 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5; 61.2% on the internal Kimi Code Bench 3.0 Long-horizon versus 55.95% for K3; 92.7% on BenchCAD with tools, below Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%); 64.4 on OfficeQA Pro v2. On an internal hallucination stress test, the claim-level hallucination rate fell from 23.4% to 8.3%. On the xdev-bench-hard cyber eval it scores 81.0% pass@3 (90.5% for the checkpoint without data removal). No independent third-party reproduction yet.

Getting started

It is usable today: pick Kimi K3.1 Pro with the /model command in Kimi Code, or call kimi-k3.1-pro / kimi-k3.1-pro-highspeed on the Kimi API platform, where Zero Data Retention is available. K3.1 Pro costs $2 input, $10 output and $0.20 cached input per million tokens, about a third cheaper than K3 at $3/$15; Highspeed is $4/$20. Enterprise admins must enable it manually. Teams planning to self-host can wait for the weights due by Nov 4 and size deployment on inference stacks such as vLLM that already support K3.