The version number of K3.1 Pro has only been bumped up by 0.1, yet Moonshot AI is aiming to address far more ambitious questions.
On the morning of October 23, Kimi officially posted a 24-second video on X and Video Accounts.
The screen shows a retro-style vintage computer, with Kimi Code running on the display and lines of Three.js code being typed into the terminal, generating a complex 3D scene. Keen-eyed netizens quickly found the answer in the model bar at the top left—“Model: Kimi K3.1.”
Kimi’s official K3.1 teaser video | Image source: X @Kimi_Moonshot
Running cutting-edge 3D code on an antiquated machine—this stark contrast is, in itself, the message Moon’s Dark Side seeks to convey.
A little over ten hours later, just after midnight on October 24—coinciding with Programmers’ Day—Kimi K3.1 Pro officially launched, accompanied by a Highspeed version that’s two to three times faster. The company has pledged to release the full model weights by November 4.
It’s been just over three months since the previous-generation K3 was released. In that span, Moonshot has gone through more in a single quarter than most companies experience in an entire year.
“More for the same price” 0.1
First, let’s clarify what the K3.1 Pro is.
It is not an entirely new model. Moonshot AI explicitly states that the K3.1 Pro is a post-training variant of the K3, with no changes to its underlying architecture—still a mixture-of-experts design boasting 2.8 trillion total parameters and 104 billion activation parameters, still featuring a 1-million-token context window, and still natively supporting image and video inputs.
What has changed is the post-training phase. According to the official statement, “significantly increased post-training compute resources” were deployed, and large-scale federated reinforcement learning was conducted across hundreds of downstream domains. To support this, the underlying engineering components—including the reinforcement learning framework, sandbox infrastructure, and long-context reasoning—were completely rearchitected.
At that point, the same brain was retrained on how to perform the task.
The results are reflected in the rankings. On Terminal-Bench 4.0, which evaluates programming agents, K3.1 Pro scored 65.7%, surpassing GPT-6 Astra’s 58.2% and trailing Claude Opus 5.5’s 66.4% by less than one percentage point. In the previous generation, K3 achieved only 13% on this benchmark, according to measurements by the third-party firm Artificial Analysis.
Third-party comprehensive ratings also give it due recognition. On the Artificial Analysis Intelligence Index, the K3.1 Pro scored 55 points, 11 points higher than the K3, ranking fourth globally—behind only Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full. Once the equal-weighted official scores are released, it will become the highest-scoring open-weight model, narrowing the gap between open-source and closed-source models from 12 points to 3 points.
K3.1 Pro’s performance across 14 programming, science, and knowledge-work benchmarks | Image source: Kimi official blog
In addition to programming, the two new main themes this time are 3D and finance.
K3.1 Pro can drive game engines like Unity, Unreal, and Godot, and can write Blender and CAD scripts for parametric modeling. In the BenchCAD test for reconstructing 3D models from multi-view images, it achieved 92.7% using its tools, approaching the levels of Opus 5.5 and Astra. Kimi’s web version and app have also simultaneously launched “3D Make,” allowing users to directly describe their requirements and generate 3D web pages, 3D models, or even export files for 3D printers.
The 3D Make interface on kimi.com | Image source: Kimi official blog
Looking back at that teaser video, the Three.js running on the old computer’s screen was, in fact, just a preview of this main storyline.
In the financial domain, K3.1 Pro scored 64.4 on OfficeQA Pro v2, which performs multi-step numerical reasoning based on real financial statements—officially the best result to date. In a high‑difficulty hallucination test designed by the developers themselves, the factual error rate dropped from 23.4% in K3 to 8.3%—though this is a deliberately adversarial stress test and does not reflect the error rate in everyday use.
Financial Review Scores | Image source: Kimi Official Blog
The most surprising thing is the price.
K3.1 Pro’s API pricing is 13 yuan per million input tokens and 68 yuan per million output tokens, which is one-third cheaper than K3 and roughly half the price of Claude Opus 5.5. The newer, more powerful model is actually cheaper than the older one, a rarity in the large-model industry.
However, Artificial Analysis also did the math: the K3.1 Pro uses 14% more output tokens to complete the same task, so the actual cost per task is about $1.98, roughly on par with the K3.
K3.1 in the real world
Beyond the rankings, first-hand user feedback is far more complex.
Positive reviews focus on its “work performance.” On Zhihu, a developer tested it with a project that mixes Rust, C, assembly, and CMake, concluding that K3.1 Pro’s coding capabilities are significantly stronger than DeepSeek V4.1 Flash—particularly in mastering common Rust design patterns—and even fixed an issue that DeepSeek had left unresolved just days earlier. For the same task, K3.1 Pro took 20 minutes, whereas switching to K3 would likely take 40 minutes to an hour.
Speed is another improvement that has been repeatedly highlighted. K3’s slowness was once notorious—some even joked it was so slow that “you could read its English thought process in a smooth, unhurried way.” By the time we reached K3.1 Pro, most of the thought process now flashes by so quickly that you don’t even have time to see it before it gets collapsed.
But front-end aesthetics have become the focal point of controversy.
A forum user conducted a single-file HTML test using their own proprietary prompts: “Cloud Dragon City,” “Iridescent Mechanical Peacock,” and “Diamond.” The conclusion was rather blunt: K3.1 Pro ranked near the bottom among the models tested, with the peacock’s highlights even rendered behind the feathers. Another developer had it recreate a Steam controller based on a reference photo using Three.js; after three iterations, the spatial relationships between the logo and the components still weren’t correct.
The same prompt, “Glazed Porcelain Mechanical Peacock,” as generated by the K3.1 Pro (Max inference) | Image source: PoemFarAway, a forum user on the Dalao Shuo community
GPT-6.1 Sol (xhigh reasoning) output under the same prompt | Image source: Forum user PoemFarAway from “Da Lao Shuo”
Reference image provided by the Steam Controller restoration test | Image source: NightRain, a forum user on the “Da Laoshuo” forum
K3.1 Pro Version 2 released; issues remain with the logo and component relationships | Image source: forum user NightRain from “Da Lang Shuo”
On one hand, there’s the stunning official 3D demo; on the other, the embarrassing “buyer’s show” mishaps shared by users. Both types of feedback are isolated incidents, but they’re enough to show that the K3.1 Pro isn’t quite at the point where you can “use it with your eyes closed.”
And the most concentrated complaints are not directed at the model, but at the pricing plans.
Some users have reported that, with the 199-yuan legacy plan, having K3.1 Pro review a small game consumed 13% of the 5-hour quota, whereas the same task took only 3% when handled by Opus. Others say the Highspeed version feels “really comfortable” to use, but its 5-hour quota doesn’t even last two hours. Meanwhile, API pricing has been cut by one-third, yet subscribers haven’t noticed any significant change.
By the third version of the test, the 199-yuan plan triggered the 5-hour usage limit | Image source: NightRain, a forum user on Dalao Shuo
This is actually an old issue that dates back to the K3 era. After K3’s July launch, Kimi temporarily suspended new consumer‑side subscriptions because the request volume far exceeded expectations. Computing power has always been the sword hanging over this company.
Another way to respond to the allegations
The timing of the K3.1 Pro’s release makes it hard not to recall the storm that has surrounded Kimi over the past few months.
In February this year, Anthropic publicly accused DeepSeek, Moonshot AI, and MiniMax of interacting with Claude and extracting its capabilities through a large number of fake accounts, with interactions attributed to Moonshot AI exceeding 3.4 million.
After the release of K3 in July, which delivered performance nearly on par with Claude Fable 5, White House Office of Science and Technology Policy Director Kratsios directly posted on X, claiming that Moonshot AI “distilled” Anthropic’s Fable model while developing K3. U.S. Treasury Secretary Janet Yellen subsequently stated that sanctions might be imposed, potentially placing it on the Entity List.
In September, Anthropic released another threat intelligence report, stating that Moonshot, through a proxy network consisting of 5,380 fake accounts, forwarded nearly 300,000 Kimi user requests to Claude over ten days and extracted the chains of thought for training purposes.
In response to these allegations, Moonlight Shadow has remained very restrained.
In July, Huang Zhenxin, head of Kimi’s enterprise business, stated that K3’s performance leap stems from original architectural innovations at the foundational level, rather than distillation or replication. By late July, while releasing K3’s model weights, Kimi also published a technical report, openly sharing its self-developed architectures—such as hybrid linear attention and attention residuals—for all to review.
So the K3.1 Pro can be seen as a silent response.
The most powerful rebuttal is never a statement, but rather a score that your opponent cannot pretend to ignore.
This time, several of the most impressive figures for the K3.1 Pro come from third parties like Artificial Analysis, rather than solely from its own benchmark tests. Even more intriguing is that, according to Anthropic’s report, Claude no longer returns a full chain of thought to external users; the distilled model “gets the address number, not the key.” Despite such restrictions, the K3.1 Pro still managed to boost its intelligence score by 11 points in three months, which at least demonstrates that Moonshot AI’s own post-training system truly delivers results.
There’s another detail that’s easy to overlook. In this release, Moonshot AI proactively disclosed that, in order to prevent the open-source community from advancing the state of the art in cyberattack capabilities, they excluded cybersecurity-related data from their training process. The trade-off was a drop in internal exploit‑testing accuracy, from 90.5% to 81%.
Internal vulnerability exploitation test results for xdev-bench-hard: the K3.1 Pro proactively implemented subtraction | Image source: Kimi official blog
At a time when it’s being labeled a “security threat,” proactively stripping features from its own open-source model is, in itself, a statement.
But it must also be acknowledged that a score can answer “Can you do it?” but cannot answer “Where did you come from?”
On the day K3.1 Pro was released, some netizens pointed out that its pelican SVG closely resembled the style of GPT-6 Astra, leading them to suspect OpenAI instead of Anthropic. Of course, this is just speculation on forums—after all, a single pelican can’t prove anything. But it does remind us that as long as the source of the training data remains unverifiable by external auditors, the “distillation” label will be hard to shake off completely.
The release of the November 4 weights will be the next point to watch. Once the weights are made public, researchers worldwide can use them for testing, benchmarking, and analysis—far more transparent than any statement.
The 3D code running on that old computer—Moonshot has already provided the answer. As for the more challenging question, the answer sheet has only just been submitted.
*Featured image source: The Dark Side of the Moon
This article is an original piece by GeekPark. For reprints, please contact Geekjun via WeChat at geekparkGO.