Moonshot AI releases Kimi K3.1 Pro and commits to open weights by November

Chinese artificial intelligence startup Moonshot AI has launched Kimi K3.1 Pro, a reasoning-focused model designed to rival top proprietary systems across coding, office work, and 3D design tasks. Alongside the flagship, the company also introduced K3.1 Pro Highspeed, a sibling variant claimed to run two to three times faster.

According to Moonshot AI's announcement, the new release serves as the first model in its K3.1 series and its most capable open-weight architecture to date. The company stated that it plans to publish the full model weights by November 4, 2026.

What happened

Kimi K3.1 Pro is a post-trained iteration of Moonshot AI's previous K3 model, scaled with significantly expanded post-training compute, according to the company. The model retains K3's architecture of 2.8 trillion total parameters and 104 billion active parameters, as well as a 1-million-token context window, The Decoder reported.

On the benchmark aggregator Artificial Analysis, K3.1 Pro scored 55 points on the Intelligence Index, according to Artificial Analysis. That marks an 11-point jump over K3 and lands the model in fourth place overall, The Decoder reported. This places it directly behind three proprietary frontier models: Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56). The Decoder noted that if the weights are published under an open license as planned, K3.1 Pro would become the highest-scoring open-weights model on that leaderboard.

Both the standard model and K3.1 Pro Highspeed are now accessible via Kimi.ai, the Kimi mobile and desktop applications, Kimi Code, and the Kimi API, according to Unite.AI.

Benchmark and agentic gains

Moonshot AI highlighted substantial progress in complex, multi-step agentic workflows. On the coding-agent benchmark Terminal-Bench 4.0, Moonshot reported that K3.1 Pro achieved 65.7 percent, surpassing GPT-6 Astra at 58.2 percent and narrowly trailing Claude Opus 5.5 at 66.4 percent, according to the company's blog. Its predecessor, K3, scored only 13 percent on the same test, The Decoder reported.

On the 3D design evaluation BenchCAD, which evaluates CAD code generation from multi-view renderings, Moonshot said K3.1 Pro scored 92.7 percent when using tools, compared to 96.2 percent for Claude Opus 5.5 and 95.9 percent for GPT-6 Astra, according to Unite.AI. Alongside the model launch, the lab rolled out a "3D Make" feature on Kimi.ai that enables users to preview, edit, and download 3D models ready for 3D printing, the company said.

Moonshot also reported internal improvements in factual accuracy. On an internal stress test called Kimi Hallu Bench Hard, the model's claim-level hallucination rate dropped from 23.4 percent in K3 to 8.3 percent, according to Moonshot AI. The lab cautioned that this evaluation probes difficult edge cases and does not reflect everyday error rates.

Deliberate cybersecurity limits

Moonshot AI explicitly addressed cybersecurity risks associated with releasing large open weights. The company explained in its technical disclosure that as post-training compute increased, the model's capacity for software vulnerability research expanded automatically because cyber data was embedded in mid-training corpora and reinforcement learning environments, according to The Decoder.

To mitigate misuse, Moonshot reported that it systematically filtered out cyber-related materials from both its training dataset and reinforcement learning environments prior to open-weights release, according to UA.News. On an internal benchmark replicating recently disclosed software exploits called xdev-bench-hard, an unconstrained test checkpoint solved 90.5 percent of tasks, whereas the filtered K3.1 Pro solved 81.0 percent, Unite.AI reported. Moonshot stated that open-weight models should avoid pushing the absolute frontier of offensive cyber capabilities while still offering defensive utility for auditing legacy code.

Pricing and infrastructure

For API users, Moonshot cut per-token fees by 33 percent relative to K3. Pricing for K3.1 Pro is set at $2.00 per million input tokens and $10.00 per million output tokens, down from $3.00 and $15.00 respectively, according to The Decoder. Cache hits cost $0.20 per million tokens.

However, real-world deployment costs may vary. Analysis by AIHubMix on Note.com noted that K3.1 Pro operates in an "always-on" thinking mode where internal reasoning tokens count toward billed output. According to Artificial Analysis, the model generated 150 million tokens during its benchmark evaluation run, making it significantly more verbose than the median model (87 million tokens) and averaging an evaluation cost of $2.11 per task.

What we don't know yet

While Moonshot AI committed to publishing model weights by November 4, 2026, it has not yet disclosed the specific open license or commercial usage terms under which those weights will be distributed, according to AIHubMix.

Additionally, K3.1 Pro's cyber benchmark results on xdev-bench-hard rely on Moonshot's self-reported evaluations and have not been independently verified, as noted by UA.News.