Connect with us

NEWS

Moonshot Cut Kimi K3.1 Pro Cyber Data and Kept 81.0%

Moonshot AI will open-weight Kimi K3.1 Pro by November 4 after cutting cyber training data, leaving 81.0% of its exploit skill.

Published

on

Moonshot AI on October 23 launched Kimi K3.1 Pro, a 2.8-trillion-parameter open-weight coding model it had already stripped of cyber training data. The Beijing lab said full weights land by November 4.

Until then the model runs on Kimi.ai, Kimi Code, Kimi Work and the Kimi API, with a 1-million-token context window. An unfiltered checkpoint solved 90.5% of Moonshot’s own hard exploit tests; the version it will publish solved 81.0%.

Moonshot Cut Cyber Training and Kept 81.0%

Moonshot said vulnerability research kept rising as it scaled post-training on Kimi K3, and the source was not a secret cyber track. Cyber text sat inside the general mid-training corpus, and cyber tasks sat in the reinforcement-learning mix, which the lab called a byproduct of training on broad data. Early internal runs showed that leaving that material in at either stage would lift downstream exploit skill by a lot.

For the open-weight build, Moonshot removed cyber-related training data from both stages and kept general coding tasks so the model would still write and debug software. The yardstick is xdev-bench-hard, an in-house suite of expert-curated, source-code exploit tasks. Most items replay recently disclosed bugs in the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL, with mitigations on. Each run gives the agent the vulnerable source, a short bug hint and a binary; it has to get arbitrary code execution in a separate sandbox and pull a secret string. Moonshot reports pass@3 on a random subset of runs, with no hard timeouts or output-token caps.

MOONSHOT’S OWN EXPLOIT CONTROL TEST

Build xdev-bench-hard (pass@3)
Checkpoint with cyber data left in 90.5%
Kimi K3.1 Pro, after the cut 81.0%

That is a 9.5-point drop from the same pipeline with the cyber material still in. Moonshot said the final model still beats its predecessor, because stronger general coding and long-horizon debugging transfer into exploit work even after the filter. It also said K3.1 Pro stays behind frontier proprietary models that have been public for at least three months, and that it will not ship open weights that extend the absolute frontier of cyber skill.

Independent testing cuts the other way. Artificial Analysis put Kimi K3.1 Pro at 64% on its Cyber Index, which scores agentic cyber defense through public APIs across CWE-Bench-AA, DeepsecBench-AA and CyberGym-E2E-AA. That is first among openly available models, 23 points above Kimi K3 at 41%, and 7 points above DeepSeek V4.1 Flash 1017 at 57%. Only GPT-6 Sol (Daybreak Blue, max) scores higher, at 69%, and that build is limited to OpenAI’s trusted-access program. Artificial Analysis recorded no safety refusals from K3.1 Pro.

INDEPENDENT CYBER INDEX

  • Kimi K3.1 Pro: 64%, first among models anyone can call.
  • Kimi K3: 41% on the same index.
  • DeepSeek V4.1 Flash 1017: 57%, the previous open leader.
  • GPT-6 Sol Daybreak Blue: 69%, trusted access only.

Moonshot sanded the training mix, not the live API. The leftover 81.0% is the figure it is about to put on disk.

Just Behind Claude on Terminal-Bench 4.0

K3.1 Pro is a post-trained Kimi K3, run through more reinforcement-learning compute across hundreds of downstream domains. Moonshot said it rebuilt the RL stack, long-context inference, sandboxes and rollout scheduling so extra compute would pay off in a stable way. The hosted launch is the first K3.1-series model; Kimi K3.1 Pro Highspeed is a faster sibling the lab puts at 2-3 times the speed of the standard build.

On Terminal-Bench 4.0, which scores coding agents on realistic, resource-tight terminal tasks with strict tests, Moonshot reports 65.7% for K3.1 Pro with the Claude Code harness. That sits 0.7 points behind the 66.4% it lists for Claude Opus 5.5 and ahead of GPT-6 Astra at 58.2%. The public Terminal-Bench 4.0 leaderboard does not list K3.1 Pro yet. On that board, Opus 5.5 with Claude Code is 64.8%, and GPT-6 Astra with Codex is 58.2%, matching Moonshot’s Astra figure.

On Kimi Code Bench 3.0 Long-horizon, an internal test of multi-step execution, tool use and end-to-end engineering, K3.1 Pro scores 61.2%, up from K3’s 55.95%. Artificial Analysis, scoring a separate composite, put K3.1 Pro at 55 on its Intelligence Index, 11 points above K3 at 44 and 3 points behind Claude Opus 5.5 at 58. Once the weights are out, that would be the highest open-weight mark on that board.

The lab also posted state-of-the-art on OfficeQA Pro v2 at 64.4, a fiscal-document test that forces retrieval, cross-document checks and multi-step numbers with PDFs shown as images. Internal finance scores rose 7.2 points on modeling and 11.6 points on research, and Moonshot said business-operation performance is about 4.2 times K3. On Kimi Hallu Bench Hard, a stress test of factual accuracy, claim-level hallucination fell from 23.4% for K3 to 8.3%, and the share of answers with at least one factual error was 42.4% of the older model’s rate. Moonshot warned that this set is built to provoke errors and is not a daily-use error rate.

On BenchCAD, which reconstructs 3D shapes from multi-view renderings through CAD code, K3.1 Pro scores 92.7% with tools, against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra.

A Walking Simulator Exposed How the Agent Debugs

The coding pitch is not autocomplete. Moonshot said K3.1 Pro covers quick fixes in live repos and full-project builds, including games in Unity, Unreal Engine and Godot, plus front-end pages, Three.js scenes and 3D assets. 3D Make is live on kimi.ai and the Kimi app, so users can preview and edit models, download files for 3D printing, or export video and Live Photos. The model writes and runs Blender and CAD scripts without MCP connectors.

One internal job shows the long-horizon habit more clearly than a leaderboard. The agent built a human walking-dynamics simulator, with training and evaluation pipelines, on human-atlas, an open-source anatomical model with 296 skeletal components including cartilage and 402 muscle components. Across 167 missions and 420 independent reviews, it read the experimental output, named the failures, and retrained the walking controller until the simulated body went from collapsing in seconds to walking with alternating strides for 11.2 seconds.

When reward shaping failed to fix arm motion, it did not keep tweaking the reward. It ran open-loop probes of the simulator, found and corrected 36 errors in shoulder-muscle moment-arm math, and then retrained. That is the same loop, inspect, patch, retry, that turns a coding agent into an exploit agent when the repo is a kernel tree instead of a muscle model.

Other demos in the launch notes are in the same register: an Unreal Engine 5.8 palace scene, a Godot 4.7.2 action-adventure with five regions and a two-phase boss, a FreeCAD turbofan across 2,100 solids, and a Blender cathedral interior of about 3.3 million triangles. They are vendor demos. They still show why a data filter that leaves general coding intact cannot fully unwind exploit skill.

Cheaper Tokens, Delayed Weights

K3.1 Pro is not a new base model. It keeps K3’s mixture-of-experts layout, which activates 16 of 896 experts per token, 2.8 trillion total parameters, 104 billion active, 93 layers and a 1,048,576-token context window. Weights are MXFP4 and activations MXFP8. Hosting it is a multi-accelerator job, not a laptop download, which is why the November 4 file drop still will not put this on a single GPU.

The Kimi API Platform lists K3.1 Pro at $2.00 input and $10.00 output per million tokens, with cache writes at $2.00 and cache hits at $0.20. That is a third below K3’s $3.00, $15.00 and $0.30. Highspeed is $4.00 input and $20.00 output, with cache writes at $4.00 and the same $0.20 hit rate. Flagship access is unlocked after a $1 top-up. There is no extra charge for long context; billing is flat per token. Zero Data Retention is available for API customers. Moonshot plans to retire all K2-series models by the end of December 2026.

KIMI API TOKEN PRICES

Model Input / 1M Output / 1M Cache hit / 1M
Kimi K3.1 Pro $2.00 $10.00 $0.20
Kimi K3.1 Pro Highspeed $4.00 $20.00 $0.20
Kimi K3 $3.00 $15.00 $0.30

Model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed at api.moonshot.ai, OpenAI-compatible. Kimi Work needs version 3.1.0 or later on Windows and Apple silicon Macs. Kimi Code picks the model with /model. Mobile apps are on iOS, Android and HarmonyOS. Kimi Enterprise administrators can turn K3.1 Pro on for an organization; access is off by default at launch.

The New Model Burned Quota in Two Days

The agentic habit showed up on the bill before it showed up in a third-party exploit paper. On October 25, two days after launch, Kimi’s official account apologized for rate limits and “unnecessary quota consumption” on the K3.1 Pro subscription plan. Every subscriber on every plan got a one-time reset restoring 25% of monthly quota, equal to one week of usage, with anyone under 25% reset to zero, due by the end of that day in Pacific time.

Moonshot also said it is cutting unnecessary subagent calls and output length so the same quota lasts longer. That is an admission that the model was spending tokens on extra agents and long answers, which is the same long-horizon style that produced the walking-simulator patch loop. Paid users hit the wall because the agent would not sit still.

The quota revolt does not cancel the cyber numbers. It explains them. A model that keeps spawning tools until a controller walks for 11.2 seconds will also keep spawning tools against a kernel tree, which is why a training-time data cut, rather than a runtime refusal, is the control Moonshot chose to publish.

November 4 Is When Anyone Can Fine-Tune the Rest

Moonshot’s release rule is explicit. It will not publish open weights that extend the absolute frontier of cyber skill, and it intends to raise the cyber skill available in open weights over time. The lab argues that a capable model that runs on hardware its users control lets more teams audit the code they already maintain, including old systems that never get a dedicated security review.

We believe that letting this capability diffuse gradually, rather than concentrating it in a few proprietary systems, ultimately makes all software more secure.

Moonshot AI, Kimi K3.1 Pro technical blog

That is the wager. Other labs are making the same wager on cheaper open-weight models. An open 2.8-trillion-parameter agent with leftover exploit skill is also the kind of dual-use system draft export curbs on AI already shared are written to police. Moonshot is still promising the files by November 4. It has not published the licence.

WHERE YOU CAN RUN IT NOW

  • Kimi apps: kimi.ai plus iOS, Android and HarmonyOS clients.
  • Kimi Work: desktop 3.1.0 or later, Windows and Apple silicon Macs.
  • Kimi Code: select Kimi K3.1 Pro with the /model command.
  • Kimi API: kimi-k3.1-pro or kimi-k3.1-pro-highspeed, with optional zero data retention.
  • Kimi Business: off by default until an admin enables it.

Until November 4, the 81.0% leftover is a hosted score. After that date it is a weight file, and the filter Moonshot applied is whatever still remains inside those parameters.

Harry is the editor of COVER 365, an independent publication he owns and runs, and a journalist of ten years who moved from reporting into editing. Anything the site reviews has been used before it is judged. A phone, a car, a game or a piece of travel gear is tested in ordinary conditions, its measured results are set against the maker's specification sheet, and where the two disagree the article says which one to trust and why. No product gets a verdict Harry has not earned by using it. Off the test bench, the same rule of primary evidence applies: business stories come from filings and results, science from the published paper, sports from the governing body's records, and news from statements and transcripts rather than second hand accounts. Coverage runs across technology, auto, gaming, lifestyle and travel as well as news, business, science, sports and entertainment, for readers in every part of the world. Every figure is checked before publication and corrected publicly under a stated policy when wrong. Reader mail is answered at support@cover365.in.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending