Kimi K3.1 Pro is live, with open weights due by 4 November

The blog says K3.1 Pro is a post-trained version of Kimi K3, with substantially larger post-training compute, including joint reinforcement learning across hundreds of domains[1]. The API site describes K3 as a 2.8-trillion-parameter model. It describes K3.1 Pro as 3T-class, with native visual understanding, and it does not print a separate parameter count for K3.1 Pro[2]. Artificial Analysis lists K3.1 Pro as 2.8 trillion total parameters and 104 billion active, taking text and image input[5]. That size is Artificial Analysis's listing. It is not a sentence in the tech blog[1].

On Terminal-Bench 4.0, Moonshot says K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%[1]. A footnote says the K3.1 Pro run used the Claude Code harness, and that the other scores are the maximum value reported in a lab's blog or on the official Terminal-Bench 4.0 leaderboard. Unless the post says otherwise, its K3.1 Pro results use reasoning effort set to max, temperature 1.0 and top-p 0.95[1].

Artificial Analysis says K3.1 Pro scores 67% on Terminal-Bench 4.0, up from 13% for Kimi K3, ahead of Claude Sonnet 5.5 at 64%, and the highest score of any model on that test[5]. The note does not mention the Opus 5.5 figure in Moonshot's post, and it does not describe a harness. Moonshot's printed score for its own model is 65.7%[1]. These are two reported measurements.

On Artificial Analysis's Intelligence Index, K3.1 Pro scores 55, an 11-point gain on Kimi K3 at 44. It trails Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Muse Wonder Full at 56. The note marks the Opus and Sonnet settings as max with fallback, and Muse Wonder Full as max[5]. Artificial Analysis says that once the weights are released, K3.1 Pro would be the highest-scoring open-weights model on that index, and the gap between the leading open-weights model and the leading proprietary model would narrow from 12 points to 3[5]. Moonshot's post still has the weights ahead of it, on 4 November 2026[1].

The same Artificial Analysis note gives further scores: third on GDPval-AA at 1,846 Elo, up 312 Elo from Kimi K3, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857; AutomationBench-AA from 58% to 73%; Humanity's Last Exam from 47% to 52%; CritPt from 23% to 25%; SciCode from 59% to 61%. Its AA-Omniscience Index rises from 20 to 30, with the hallucination rate on that index falling from 53% to 44% and accuracy rising from 48% to 51%[5]. Moonshot's footnote says its own GDPval-AA, AA-Briefcase and AutomationBench-AA numbers are cited from Artificial Analysis[1].

Kimi Code Bench 3.0 Long-horizon, which Moonshot describes as an internal test of multi-step coding, is 61.2% for K3.1 Pro and 55.95% for K3[1]. On BenchCAD with tools, the post says 92.7%, approaching Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. The chart on the page is labeled with a without-tools run: K3.1 Pro at 72.8 and Opus 5.5 at 73.0, with Claude Fable 5.1 at 60.6 and DeepSeek-V4.1-Flash-1017 at 41.2[1]. A footnote says the open models on BenchCAD were run by Moonshot on a 1,000-case subset of a 17,900-case set, and that the Opus 5.5 and Fable 5.1 scores are taken from the Claude Opus 5.5 system card[1].

On OfficeQA Pro v2, the post says 64.4 and calls the result state of the art. The finance chart on the page labels Frontier Finance at 58.6. Against K3, the internal chart labels Financial Modeling at 57.6 versus 50.4, Financial Research at 70.4 versus 58.8, and Business Operation at 37 versus 8.9. The prose calls those gains 7.2 points, 11.6 points, and about 4.2 times K3[1].

The post also describes one internal coding run: a walking-dynamics simulator built on human-atlas, which it calls an open-source anatomical model with 296 skeletal components and 402 muscle components. Across 167 missions and 420 independent reviews, it says the simulated human went from collapsing within seconds to walking with alternating strides for 11.2 seconds, after the model corrected 36 errors in shoulder-muscle moment-arm calculations[1]. That is Moonshot's account of the run.

The API guide's paragraph on factual accuracy describes an evaluation of de-identified conversations in which users had flagged factual errors. It says the claim-level hallucination rate fell from 23.35% for K3 to 6.85%, and that the share of responses with at least one factual error was 46.4% of the predecessor's rate[3]. The tech blog's Kimi Hallu Bench Hard paragraph says the claim-level rate fell from 23.4% for K3 to 8.3%, and that the share of responses with at least one factual error was 42.4% of K3's rate. It says the set is a stress test and does not represent everyday error rates[1]. The hallucination chart on the blog is labeled with a without-tools rate of 5.4% for K3.1 Pro, against 4.2% for GPT-6 Astra, and a hard-with-tools rate of 8.3%, against 5.5% for Astra[1]. The 6.85% figure and the 8.3% figure are on different Moonshot pages. This story does not merge them.

The API guide's Terminal-Bench 4.0 sentence still has bracketed placeholders where a score would go. The 65.7% in this story is the figure printed in the tech blog[3][1]. The API guide says K3.1 Pro achieves state-of-the-art results on SpreadsheetBench v2 and does not print a score there. The tech blog does not mention SpreadsheetBench[3][1].

On the API site, kimi-k3.1-pro is US$2.00 per million input tokens, US$10.00 per million output tokens, US$2.00 per million cache writes and US$0.20 per million cache hits. K3 on the same page is US$3.00, US$15.00, US$3.00 and US$0.30[2]. Artificial Analysis describes the input, output and cache-hit rates as US$2, US$10 and US$0.20, 33% lower than K3 on all three, and says a typical Intelligence Index task costs about US$1.98 because K3.1 Pro uses 55,000 output tokens per task, 14% more than K3's 48,000[5].

Highspeed is listed at US$4.00 input, US$20.00 output and US$4.00 cache writes, with the cache hit still at US$0.20[2]. The API site calls it the same intelligence at roughly 3× the speed[2]. The tech blog says 2–3×[1]. The pricing guide says every K3-series model supports a 1,048,576-token context window, and that context-cache billing is on by default with a 5-minute time to live. Cache writes for the series are billed separately for a 5-minute or a 1-hour time to live[4].

The API guide says reasoning is always on. The request field is reasoning_effort, with low, high and max, and max is the default[3]. The blog's product list is the Kimi apps on iOS, Android and HarmonyOS, Kimi Work version 3.1.0 or later for Windows and Apple silicon Macs, Kimi Code through the /model command, and the API ids kimi-k3.1-pro and kimi-k3.1-pro-highspeed. It says Zero Data Retention is available for API customers. For Kimi Enterprise, administrators can enable the model, and access is off by default at launch[1]. The post also says 3D Make is live on kimi.ai and the Kimi app, for previewing and editing models in the browser[1].

Moonshot says that for the open-weight release it identified and removed cyber-related data from mid-training and from the reinforcement-learning mix, and that it kept the general-purpose coding tasks[1]. On xdev-bench-hard, its in-house benchmark, it says the checkpoint without those removals solves 90.5% of problems at pass@3 and the final K3.1 Pro reaches 81.0%. It says the final model still improves substantially over K3, and that its offensive cyber capability remains behind frontier proprietary models that have been public for at least three months[1]. The chart on the post labels Kimi K3 at 33.3 on that bench[1].

The post says the company aims not to release open weights that would extend the absolute frontier of cyber capability, and that it intends to raise the cyber capability available in open weights over time[1]. The weights themselves are not in the post. Artificial Analysis says they are not published[1][5].

K3.1 Pro is on sale before the file that would make it an open-weight model exists. Moonshot's account is a post-train of K3 that, on its own terminal-coding run, sits just behind the Opus 5.5 score it cites, at a listed input, output and cache-hit price below K3[1][2]. Artificial Analysis's account is a fourth-place Intelligence Index score, a Terminal-Bench figure of 67%, and an open-weights lead that it places after the weights ship[5]. The API guide's benchmark paragraph still has placeholders where the tech blog has numbers, and the two Moonshot pages do not share one hallucination rate[3][1]. The date Moonshot prints for the weights is 4 November 2026[1].