In the early hours of October 24, Moonshot AI released its next-generation open-source model, Kimi K3.1 Pro.

It is the first model in the K3.1 series, with 2.8T total parameters, 104 billion active MoE (mixture-of-experts) parameters and support for a 1 million-token context window. By the company's own account, it retains the title of the world's largest open-source model. Full model weights will be released by November 4.

Following the launch, K3.1 Pro ranked fourth globally on the Artificial Analysis intelligence index and first among open-source models worldwide, scoring 11 points higher than the previous-generation Kimi K3.

A Post-Training Version That Keeps the Largest Open-Source Model Crown

K3.1 Pro is a post-trained version of Kimi K3. Moonshot AI says it sharply increased post-training computing power and extended RL (reinforcement learning) training across hundreds of downstream domains. To do that, it rebuilt its RL framework, long-context inference, training efficiency, sandbox infrastructure and scheduling system, with the goal of converting every additional unit of computing power into a stable return.

That also explains a detail in the official wording: this generation did not change the parameter count, which is why the company says it "retains" rather than "resets" the record. The gains come from pouring more computing power into post-training, not from stacking on more parameters.

On Terminal-Bench 4.0, which tests coding agents, K3.1 Pro scored 65.7%, beating GPT-6 Astra's 58.2% and coming within a percentage point of Claude Opus 5.5's 66.4%. On Kimi's internal long-horizon coding benchmark, K3.1 Pro scored 61.2%, up from K3's 55.95%.

On coding execution, the company says K3.1 Pro can build a complete project from scratch. It showcased the model constructing an Eastern mythological celestial palace above a sea of clouds in Unreal Engine 5.8, and building from scratch in Godot 4.7.2 a 3D action-adventure game with five exploration zones and a two-phase boss fight. Both cases were delivered end to end.

3D Scores Close In on Claude, but Rendering Is Not Its Job

3D modeling is another focus of this upgrade.

On BenchCAD, a benchmark that measures the ability to generate CAD (computer-aided design) code and reconstruct 3D models from multi-view renderings, K3.1 Pro scored 92.7% with tool use, close to Claude Opus 5.5's 96.2% and GPT-6 Astra's 95.9%. According to the company, give K3.1 Pro a text description, a sketch or a reference image and it can infer an object's shape, proportions and spatial relationships, then reconstruct it as a usable 3D model.

This is not Kimi's first foray into 3D. After K3 launched in July, users spotted a web game platform called K399 where every game was built with help from the K3 model, spanning shooters, rhythm games, 3D puzzles and other genres. Most were recreations of established titles, with a smaller share of original work. The company now says K3.1 Pro is adept with game engines including Unity, Unreal Engine and Godot, and can combine web search with 3D generation to deliver a complete experience.

Seen against the wider industry, this field already has earlier entrants. In April, ByteDance released Seed3D 2.0, a next-generation 3D generation model that the company says hits SOTA (state of the art) on both geometric precision and material generation, with support for part-level decomposition, articulated asset generation and scene composition. Around the same time, Tencent released and open-sourced Hunyuan 3D World Model 2.0 (HY-World 2.0), which accepts text, image and video input and can automatically generate, reconstruct and simulate complete 3D worlds. In August, Tencent followed with WorldClaw, a 3D open-world generation framework in which a single open-ended text prompt produces a large-scale 3D world that users can freely explore and keep editing.

AI 3D players such as Deemos Technology and VAST have also raised funding, with investors and tech giants piling in during the same window.

High scores, though, do not mean full capability. Guo Tao, an angel investor and veteran AI expert, says K3.1 Pro leans on code and spatial reasoning to directly output scripts that modeling software and game engines can run, completing parametric asset construction. But it lacks the underlying capability for independent real-time rendering and high-quality texture generation, and on the fineness of art assets it still trails specialist 3D vendors and the native world models from ByteDance and Tencent.

Put differently, what K3.1 Pro does well is translate ideas into executable modeling instructions, not draw the picture itself.

Finance and Law, the Two Scenarios Post-Training Leans Into

Having entered its "3.0 era," Moonshot AI has shifted its focus to delivery within specific scenarios.

On September 17, Kimi launched a solution for the financial industry, pulling together more than 10 authoritative data sources, nine financial skills, five compliance and security measures, and institutional-grade data modeling and report delivery. Dozens of financial institutions are running business innovation projects or co-building solutions on Kimi's models, enterprise edition and Kimi Work, including CICC, China Securities, Sequoia China, Ant Group and Meituan.

With K3.1 Pro, finance is again one of the scenarios post-training leans into, and the company says the model can deliver complete financial workflows end to end. On external benchmarks, K3.1 Pro scored 64.4 on OfficeQA Pro v2, a SOTA result, with dedicated optimization for the factual accuracy that financial work demands. On the internal stress test Kimi Hallu Bench Hard, the new model's error rate on individual factual statements fell from K3's 23.4% to 8.3%, and the share of answers containing at least one factual error was roughly 42.4% of the previous generation's.

Law is the other main thread. A post from Moonshot AI in late August showed that on Harvey LAB, the benchmark closest to real law firm work, Kimi K3 improved from K2.6's score of zero to 26.7%. Harvey, the legal AI company behind Harvey LAB, subsequently released Tenet, its first legal-industry model built on Kimi K3.

Kimi is not alone in this choice. In June, OpenAI rolled out six Codex role plugins at once, covering equity investing, investment banking, data analysis, creative production and other jobs, and teased plugins for law and corporate finance. In August, Alibaba's Qwen3.8-Max arrived with a 2.4 trillion-parameter MoE architecture aimed at professional tasks in law, finance and design, supporting million-token context and long-horizon autonomous planning. Scenario specialization has become the consensus route for leading model companies competing for enterprise spending.

In Guo's view, Kimi's move into finance and law builds on its own strength in long-document comprehension rather than chasing a trend. He also notes that few domestic foundation models combine comparable long-text capability with on-premises deployment options, giving Kimi a chance to accumulate industry workflows and tool chains and build a first-mover advantage, at least for a while. But high-rigor work still requires human final review, and willingness to pay is constrained by compliance rules: standardized support processes can be monetized reliably, while core decision-making is unlikely to be fully handed to a model any time soon.

Distribution Reaches Bedrock and Codex

Beyond scenario deployment, the pace on funding and distribution has been just as tight.

Reports in early October said Moonshot AI had closed its final pre-IPO private funding round at a valuation of about $50 billion and was moving toward a Hong Kong listing in the first quarter of next year. Back in July, its valuation had already climbed from $10 billion to $35 billion post-Series F over the course of the year.

Beyond the listing timetable, Kimi's strategy extends past scenario deployment: it also wants a cut of global cloud providers' business through revenue sharing.

On September 18, Amazon Bedrock announced it was adding Kimi K3, letting enterprise developers worldwide call the model directly. It was the first time a Chinese large model company entered a global cloud provider's model catalog on a revenue-sharing basis. In late September, US AI infrastructure company Baseten said enterprise users could use Kimi K3 inside OpenAI's coding tool Codex, with usage billed directly against existing OpenAI purchase commitments and no new procurement process required. That marked the first time a Chinese open-source model entered the mainstream paid settlement channel for OpenAI's enterprise customers.

The domestic market is seeing parallel rollouts. On October 24, Tencent's WorkBuddy announced it had added Kimi K3.1 Pro, along with K3.1 Pro Highspeed, a faster version with equivalent intelligence and roughly two to three times the output speed, initially available to enterprise flagship and individual subscription users.

On the Eve of an IPO, Computing Power Remains the Hardest Constraint

The heavier the scenario, the sooner the cost shows up.

Guo notes that pushing deeper into compute-heavy scenarios such as finance, law and AI 3D will keep pressure on computing power supply. These multi-step tasks, with repeated tool calls, consume far more resources than ordinary chat. As usage from enterprise customers and 3D creators rises, cluster scheduling and compute reserves will become a bottleneck on expansion, and the company will need caching, scheduling strategies and tiered computing power plans to balance cost against service stability.

This is not hypothetical. Within 48 hours of K3's launch, Kimi posted that user request volumes had far exceeded its projections and were approaching the limits of its existing clusters, and said it was pausing new consumer subscriptions.

Having pivoted from the parameter race to scenario delivery, Moonshot AI is accelerating on three fronts at once: technology, scenarios and revenue. Whether it can turn scenario delivery into a sustainable revenue curve remains the question this unicorn has to keep answering on the eve of its listing.