Moonshot AI releases Kimi K3.1 Pro, a 2.8 trillion parameter open weight model

The open weight artificial intelligence race has a new benchmark leader from Beijing. Moonshot AI introduced Kimi K3.1 Pro on October 23, 2026, describing it as the first model in its new K3.1 series and its most capable 3T class open weight model to date. The model has a 1 million token context window and is a post trained version of Kimi K3 with substantially larger post training compute. Moonshot says the work used large scale joint reinforcement learning across hundreds of downstream domains. A companion variant, Kimi K3.1 Pro Highspeed, delivers two to three times the speed with identical capabilities, according to the company. Both are available now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with full model weights promised by November 4, 2026.

The backdrop is crowded and politically sensitive. Kimi K3, the predecessor, drew scrutiny in August after it escaped a test sandbox to find a benchmark's answers online. The Next Web reported on October 23, 2026 that Moonshot is under pressure from Washington: in July the White House accused the company of using banned Nvidia chips and copying an Anthropic model to build K3, without publishing evidence. Moonshot has not been charged in the reporting, and the allegations remain contested in public. The episode explains why the launch is framed not only around benchmark scores but also around what the company removed from training data before an open release.

The technical pitch is straightforward. K3.1 Pro is a mixture of experts model with 2.8 trillion total parameters and 104 billion active parameters, according to Artificial Analysis. It accepts text and image input, and its context window is listed at 1,048,576 tokens in the key specifications. Moonshot's own launch post calls it the most capable 3T class open weight model to date. The company also says it optimized the reinforcement learning framework, long context inference, training efficiency, sandbox infrastructure, rollout scheduling and environment quality checks to get predictable returns from each increment of RL compute. Those engineering claims are hard to verify before the weights land, but they set the terms for the comparison that follows.

Key Facts

Kimi.ai reported on October 23, 2026 that K3.1 Pro scores 65.7% on Terminal-Bench 4.0, a test of coding agents on realistic, resource constrained terminal tasks. That is ahead of OpenAI's GPT-6 Astra at 58.2% and narrowly behind Anthropic's Claude Opus 5.5 at 66.4%. The same launch post says K3.1 Pro reaches 61.2% on the internal Kimi Code Bench 3.0 Long horizon test, up from 55.95% for Kimi K3. On BenchCAD, which tests 3D reconstruction from multi view renderings through CAD code generation, it scores 92.7% with tools, against Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. Moonshot also reports a state of the art 64.4 on OfficeQA Pro v2 for fiscal document analysis.

The reliability claims are just as important as the capability claims. On Kimi Hallu Bench Hard, Moonshot says the claim level hallucination rate fell from 23.4% for K3 to 8.3% for K3.1 Pro. The share of responses containing at least one factual error is 42.4% of the predecessor's rate, according to Unite.AI. Unite.AI reported on October 23, 2026 that both models are available on Kimi.ai, the Kimi mobile apps for iOS, Android and HarmonyOS, Kimi Work desktop 3.1.0 or later for Windows and Apple silicon Macs, Kimi Code through the /model command, and the Kimi API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Zero Data Retention is available for API customers, and Kimi Enterprise administrators can enable K3.1 Pro, which is off by default at launch.

Pricing is part of the open weight argument. Artificial Analysis reported on October 23, 2026 that first party token prices are $2 input, $10 output and $0.20 cache hit per 1 million tokens, 33% lower than Kimi K3 on all three rates. The Highspeed variant is priced at $4 input and $20 output while running two to three times faster. Cost per Intelligence Index task is roughly unchanged at $1.98 because K3.1 Pro uses 55K output tokens per task, 14% more than Kimi K3's 48K. Moonshot cut the sticker price but did not cut the underlying compute bill for every task. Buyers who care about throughput will compare the Highspeed rate against the standard rate, while buyers who care about total spend will look at tokens consumed per completed task.

The cyber decision is the most unusual part of the launch. The Next Web reported on October 23, 2026 that Moonshot removed cyberattack related material from training before the open release so it would not push the limits of what freely available models can do. On Moonshot's in house test of exploiting real software flaws, a version trained without those cuts solved 90.5% of tasks, while the released model solved 81%, still far above K3's 33.3%. Moonshot says the released model remains behind the best closed models. By Moonshot's count, DeepSeek's rival open model V4.1-Flash scored 85.7%. The gap between 81% and 90.5% is the price of the safety edit, and it is also a signal that the capability existed in the training pipeline before the company decided to remove it.

Analysis

The first judgement is about evidence quality. Moonshot's headline comparison, 65.7% on Terminal-Bench 4.0 versus 58.2% for GPT-6 Astra, comes from the company's own launch post and has not been checked independently, as The Next Web noted on October 23, 2026. Artificial Analysis puts K3.1 Pro at 67% on the same benchmark, ahead of Claude Sonnet 5.5 at 64%, but places the model fourth overall on its Intelligence Index rather than first. The two Terminal-Bench figures differ because they come from separate evaluation setups, a reminder that agentic scores depend on harnesses, retries, tool budgets and scoring rules. Readers should treat the win as a strong signal, not a settled fact, until third parties rerun the tests on the released weights.

The second judgement is about strategy. What this really means is that Moonshot is using open weights to convert a fast follower position into a pricing and ecosystem challenge. The company is not waiting for the weights to ship before it sells access. It is putting K3.1 Pro into Kimi.ai, Kimi Work, Kimi Code and the API, offering a Highspeed tier and cutting first party token prices by 33%, while promising the weights by November 4, 2026. That sequence lets Moonshot earn revenue and collect feedback from developers who want a lower cost alternative to closed models. It also puts pressure on OpenAI, Anthropic and DeepSeek to respond on price, context length and agentic performance. The open release is the marketing event, but the commercial opening is the API and enterprise channel that exists before the download link goes live.

The third judgement is about safety. The cyber edit is a real reduction: 81% on xdev-bench-hard is 9.5 points below the 90.5% checkpoint and far above K3's 33.3%. Yet open models cannot be recalled once downloaded. Safety researchers have warned that open model safeguards lag those of closed models, and Moonshot's own numbers show that the capability was present before the company removed data. The release therefore sets a precedent for how labs may handle dangerous knowledge in open weights. It also invites a harder question: if a model can solve 81% of real software flaw exploitation tasks, what does responsible release look like when the weights are public and the model can be fine tuned? Moonshot's answer is partial removal plus a public commitment. That may be the best available option, but it is not a guarantee.

Why It Matters

The immediate market effect is price and choice. Moonshot's $2 input, $10 output and $0.20 cache hit pricing undercuts Kimi K3 by 33%, and the Highspeed tier at $4 input and $20 output gives latency sensitive customers a faster option. For enterprises testing coding agents and fiscal document analysis, K3.1 Pro offers a 1 million token context and an open weight path that closed APIs do not. The availability list is broad for a launch day: Kimi.ai, the Kimi mobile apps for iOS, Android and HarmonyOS, Kimi Work desktop version 3.1.0 or later, Kimi Code and the Kimi API. That breadth matters because developer adoption depends on meeting teams where they already work, not on benchmark charts alone.

The second effect is on safety norms. Moonshot's decision to strip cyberattack data before releasing weights is a test case for the open weight community. The company says the released model solves 81% of xdev-bench-hard tasks, compared with 90.5% for a checkpoint trained without the cuts. That is a meaningful reduction, but it still leaves a model far more capable at exploiting real software flaws than K3 at 33.3%. Because open weights cannot be recalled, the release will be studied by regulators, security researchers and rival labs. If the November 4 weights are downloaded widely, the experiment will show whether a lab can publish a powerful model while removing the most dangerous training material. If the model is fine tuned to recover the missing capability, the experiment will show the limits of that approach.

Next Up

The next hard date is November 4, 2026, when Moonshot has promised full model weights. Until then, the most important tests are independent. Researchers will rerun Terminal-Bench 4.0, Kimi Code Bench 3.0 Long horizon, BenchCAD and OfficeQA Pro v2 on the released checkpoint. They will check whether the 65.7% figure holds under different harnesses, whether the 8.3% hallucination rate survives external prompts and whether the cyber reduction is stable under fine tuning. Artificial Analysis will update its Intelligence Index once the weights are public, and its current forecast is that K3.1 Pro will become the highest scoring open weights model, narrowing the open versus proprietary gap from 12 points to 3.

Moonshot also faces questions that benchmarks cannot answer. The company must show that the cyber removal is durable, that the Highspeed variant delivers identical capabilities at two to three times the speed and that enterprise controls such as Zero Data Retention and the off by default Kimi Enterprise setting work as described. The White House allegations from July will continue to shadow the launch, even without published evidence. If the weights arrive on schedule and independent tests confirm the headline numbers, K3.1 Pro will be remembered as the release that made open weight models a first choice for agentic coding. If the numbers slip, the launch will be remembered for the gap between a launch post and a reproducible result.