Tech
Moonshot releases Kimi K3.1 Pro at 3T parameters with 1M context and plans weights for November 4.
Moonshot AI (via The Decoder) · 2026-10-23 · Source ↗
Moonshot releases Kimi K3.1 Pro at 3T parameters with 1M context and plans weights for November 4.
Moonshot AI (via The Decoder) · 2026-10-23 · Source ↗
Moonshot AI announced Kimi K3.1 Pro as a post-trained extension of Kimi K3, built with substantially more post-training compute and sized at roughly 2.8 trillion total parameters with 104 billion active per step. The context window runs to 1 million tokens and the model accepts text and images. A companion variant, Kimi K3.1 Pro Highspeed, is described as 2-3x faster than the base model. Moonshot committed to releasing the full open weights by November 4, 2026, through its Hugging Face and GitHub channels.
Benchmarks against Opus 5.5
Moonshot's own table, run at reasoning effort "max," reports Terminal-Bench 4.0 at 65.7 against Claude Opus 5.5 at 66.4 and GPT-6 Astra at 58.2; SWE Marathon v1.1 at 51.7 against Opus 5.5 at 46.3; BrowseComp 92.1 against 90.8; GPQA-Diamond 94.2 against 93.9; and OfficeQA Pro v2 at 64.4, described in the post as state of the art. On the Artificial Analysis Intelligence Index, which uses a different evaluation protocol, K3.1 Pro lands at 55 behind Opus 5.5 at 58, with Sonnet 5.5 and Muse Wonder Full tied at 56. ExploitBench is the outlier at 74.7 against Opus 5.5's 91.0.
Hallucination drop and the cyber-data removal
On Moonshot's internal Kimi Hallu Bench Hard, claim-level hallucinations fell from 23.4 percent on K3 to 8.3 percent on K3.1 Pro with tools, 5.4 percent without tools, against GPT-6 Astra at 4.2 percent and Opus 5.5 at 6.8 percent. Moonshot states that for the open release it stripped cyber-related training data from mid-training and reinforcement learning, which dropped xdev-bench-hard from a 90.5 percent checkpoint to the released 81.0 percent. "No open-weights model should push the absolute frontier of offensive cyber capabilities," the post says.
"No open-weights model should push the absolute frontier of offensive cyber capabilities." — Moonshot AI, Kimi K3.1 Pro release post, 2026-10-23
Pricing and specific capabilities
Moonshot lowered API pricing with the launch: input $2 per million tokens (from $3), output $10 (from $15), cache hits $0.20 (from $0.30), a roughly 33 percent cut. K3.1 Pro Highspeed runs $4 input, $20 output, $0.20 cache. Each Intelligence Index task costs about $1.98 end-to-end, roughly the same as K3. The post highlights long-horizon RL work on an internal human-walking simulator built on an open-source human-atlas model, in which the model took a controller from collapsing within seconds to walking for 11.2 seconds after finding and fixing 36 shoulder-muscle moment-arm errors. Kimi Enterprise admins have to opt in; the model is off by default.