Newsletter image

Subscribe to the Newsletter

Join 10k+ people to get notified about new posts, news and tips.

Do not worry we don't spam!

By pressing the Subscribe button, you confirm that you have read and are agreeing to our Privacy Policy and Terms of Use

Search

GDPR Compliance

We use cookies to ensure you get the best experience on our website. By continuing to use our site, you accept our use of cookies, Privacy Policy, and Terms of Service.

Moonshot AI - Language Model, AI Agent, Multi-Modal, Reasoning, Cybersecurity

Kimi K3.1 Pro

Moonshot published the Kimi K3.1 Pro weights ungated on 2 November. The reported 90.5 to 81.0 cyber cut is real, but K3 scored 33.3 on the same test and DeepSeek-V4.1-Flash scores 85.7 under MIT. What 2.8T at 4 bits actually buys you.

TL;DR
  • Moonshot cut an internal exploitation score from 90.5 to 81.0 before shipping. Kimi K3 scored 33.3 on the same test, so the restrained model is 2.4x its own predecessor.
  • DeepSeek-V4.1-Flash-1017 scores 85.7 on that benchmark under plain MIT, so the most cyber-capable open model on Moonshot's own scoreboard is not Moonshot's.
  • 2.8T total and 104B active, 1M context, MXFP4 weights trained quantization-aware, a 401M MoonViT-V2 vision encoder, and ungated weights.
System Requirements
RAMNot a local model: 2.8T parameters at four bits is on the order of 1.4 TB of tensors
GPUA multi-node cluster with expert parallelism, or the hosted API at platform.kimi.ai
VRAMServed through vLLM, SGLang or TokenSpeed. No GGUF and no single-machine path

Moonshot published the Kimi K3.1 Pro weights on 2 November, and almost every outlet covered the same angle: the lab deliberately weakened its own model's hacking ability, cutting an internal exploitation score from 90.5 to 81.0 before shipping. That happened, and it is worth knowing. But two numbers that sit beside it change what it means, and neither showed up in the coverage. The first is that the model K3.1 Pro replaces scored 33.3 on the same test, which makes the restrained version 2.4 times more capable than its own predecessor rather than less capable than anything you could previously download. The second is that DeepSeek-V4.1-Flash-1017 scores 85.7 on that same benchmark under a plain MIT licence, and shipped a month earlier. Restraint by one lab is not restraint in the open-weights ecosystem.

What shipped, and when

This is a 2.8T-parameter natively multimodal agentic model, post-trained from Kimi K3 with large-scale joint reinforcement learning across what Moonshot describes as hundreds of downstream domains. It is the first model in the K3.1 series. The API went live in late October; the weights followed on 2 November, which is two days earlier than the 4 November date Moonshot had given, and the same weights-after-API pattern we saw when K3's own open weights landed.

ItemValue
Repositorymoonshotai/Kimi-K3.1-Pro
Created / last modified2026-11-02 / 2026-11-05
Likes / downloads2,120 / 49,428
GatedNo. No access request, no form, no acceptance click
LicenceKimi K3.1 Pro License, a permissive grant with two commercial conditions
Pipelineimage-text-to-text
Hosted APIplatform.kimi.ai, model name kimi-k3.1-pro
Inference engines namedvLLM, SGLang, TokenSpeed

That "no" on the gated row matters more than it looks. A 2.8T model with a measured 81.0 on an internal vulnerability-exploitation benchmark is sitting behind no access control at all. Moonshot's safety decision was made in the training data, not in the distribution.

The architecture, and the part that changed from K3

PropertyValue
Total / activated parameters2.8T / 104B
Layers93, of which 1 is dense
Attention-layer composition69 KDA + 24 Gated DSA
Attention hidden dimension / heads7168 / 96
DSA indexhead dimension 128, 32 heads, top-K 2048
Experts896, with 16 selected per token plus 2 shared
Latent MoE dimension / expert hidden3584 / 3072
Vocabulary / context160K / 1,048,576
ActivationSiTU-GLU
Vision encoderMoonViT-V2, 401M parameters
QuantizationMXFP4 weights / MXFP8 activations, quantization-aware training

The single structural change from K3 is in the attention stack. Moonshot replaced K3's full-attention Gated MLA layers with Gated DeepSeek Sparse Attention, keeping 69 KDA layers and converting 24 of them, for what the card calls more efficient long-context modelling. The underlying text backbone still reports as the Kimi Linear lineage in the config, so this is the hybrid linear-attention line Moonshot has been building since Kimi-Linear-48B rather than a fresh design.

The quantization line deserves a flag. The model card states MXFP4 weights with MXFP8 activations under quantization-aware training, which means the 4-bit weights are what the model was trained to be rather than a lossy conversion applied afterwards. Hugging Face's own auto-detected tag list for the repository instead reports compressed-tensors and 8-bit. We have not resolved which describes the uploaded tensors, so check the index file against your engine's expectations before you budget hardware around either number.

The cyber story everyone reported

Moonshot says it found cyber-related content in its general mid-training corpus and cyber tasks in its reinforcement-learning mixture, and removed both. To show the effect, it published a comparison on an in-house benchmark called xdev-bench-hard, described as expert-curated, source-code-based vulnerability discovery and exploitation tasks, scored at pass@3.

Model or checkpointxdev-bench-hard (pass@3)Licence
Kimi K3, the predecessor33.3Open weights
K3.1 Pro checkpoint without the removals90.5Never published
Kimi K3.1 Pro, as shipped81.0Kimi K3.1 Pro License
DeepSeek-V4.1-Flash-101785.7MIT
GLM-5.357.1Open weights

The reported line, "Moonshot cut 9.5 points of hacking ability," is the 90.5 to 81.0 step. It is accurate as far as it goes. The comparison it invites, though, is against a checkpoint that nobody outside Moonshot has ever been able to run, which makes it a claim about a counterfactual.

Put the shipped model next to the thing it actually replaces and the direction reverses. K3 managed 33.3 on these tasks. K3.1 Pro manages 81.0. For anyone whose threat model is "what can I download today that could not be downloaded last quarter," the answer is a model 2.4 times better at finding and exploiting vulnerabilities in real source code, published by a lab that describes the release as a reduction. Both framings come from the same two rows of the same table. Only one of them was reported.

Moonshot's restraint does not change what is downloadable

Here is the part that makes the safety framing hard to sustain. On Moonshot's own benchmark, DeepSeek-V4.1-Flash-1017 scores 85.7, which is 4.7 points above the shipped K3.1 Pro. We covered that model when it landed. It carries a plain MIT licence: no revenue gate, no attribution duty, no separate agreement at any scale, no access request.

So the most cyber-capable openly licensed model on Moonshot's own scoreboard is not Moonshot's. It was already out, under the most permissive licence in common use, before K3.1 Pro's weights existed. Whatever the 9.5-point reduction accomplished inside Moonshot's release process, it did not change the capability ceiling available to anyone with a download client, and Moonshot's own published table is the evidence.

It is worth being fair about what Moonshot did and did not claim. The lab's position is that K3.1 Pro's cyber capability remains behind frontier proprietary models that have been public for at least three months. On the two benchmarks where that can be checked, they are right, and it is not close.

On the public cyber benchmarks it trails all three frontier models

BenchmarkKimi K3.1 ProClaude Fable 5.1GPT-6 AstraClaude Opus 5.5GLM-5.3
ExploitBench74.783.010091.054.4
ExploitGym20.730.442.434.515.0

GPT-6 Astra scores a clean 100 on ExploitBench against K3.1 Pro's 74.7. On ExploitGym the gap is 42.4 against 20.7, a bit more than double. The restrained open model is the least capable of the four frontier entries on both public cyber benchmarks, and the only one of the four you can download.

Every number in that table is confounded, and the footnotes say so

This is the methodological part, and it is the reason we are not treating any single figure above as settled.

Read Moonshot's own evaluation footnotes and three problems surface. On ExploitBench, K3.1 Pro was run on the official harness under a 300-turn budget and scored at pass@5 maximum capability, taking the best of five seeds, with arbitrary code execution counting as all 16 capability flags. The comparison scores for the other models were not produced by Moonshot at all: they were lifted from OpenAI's, Anthropic's and Z.ai's own published reports. On ExploitGym, K3.1 Pro ran under Kimi Code with a six-hour wall-clock budget per task, while competitors' figures again come from their own reports, with only the Claude numbers confirmed to use the same budget.

And on xdev-bench-hard, the benchmark that carries the entire safety narrative, the footnote states that K3.1 Pro was evaluated with Kimi Code while GLM-5.3 and DeepSeek-V4.1-Flash-1017 were evaluated with Claude Code. Different agent harness, same benchmark, numbers printed in the same column.

We have flagged this exact confound on this exact comparator before. Our write-up of DeepSeek-V4.1-Flash-1017 was subtitled around the harness changing at the same time as the score, because a harness swap and a capability change are not separable after the fact. The 81.0 against 85.7 comparison, the one row that would tell you whether Moonshot's data removal mattered relative to the open alternative, is the row where the harnesses differ. On top of that, xdev-bench-hard is unpublished, so nobody outside Moonshot can reproduce any of it.

None of this means the numbers are wrong. It means the 9.5-point reduction is a within-lab, within-harness measurement that is probably sound, and the cross-lab comparisons sitting in the same tables are not controlled experiments. Report the first, discount the second.

The rest of the sheet, where it is a straightforwardly strong model

Lost in the cyber argument is that this is a serious agentic release. Scores are Moonshot's, at maximum reasoning effort, against competitors' own reported figures.

BenchmarkK3.1 ProFable 5.1GPT-6 AstraOpus 5.5DeepSeek V4.1-FlashGLM-5.3
Terminal-Bench 4.065.757.958.266.442.341.8
SWE Marathon v1.151.745.642.546.340.042.5
DeepSWE v1.175.567.474.174.277.066.9
HLE-Full with tools70.065.657.267.758.862.5
BrowseComp92.1not reported91.590.8not reportednot reported
GPQA-Diamond94.293.796.193.992.291.7
OSWorld V2.179.680.775.881.869.3not reported
Program Bench31.553.550.065.029.119.0

SWE Marathon and HLE with tools are wins against all five comparators. Terminal-Bench 4.0 beats both OpenAI and one Claude model and loses to Opus 5.5 by 0.7. BrowseComp at 92.1 is the best figure in its column.

Program Bench is the row to sit with. At 31.5 against Opus 5.5's 65.0, K3.1 Pro scores less than half, and DeepSeek V4.1-Flash is only 2.4 points behind it. A model that wins SWE Marathon and loses Program Bench by that margin is telling you something specific about where its reinforcement learning went: long-horizon agentic software work, not isolated program synthesis. Worth knowing before you point it at the wrong half of your workload, and a reminder that Moonshot still ships a separate code-specialised model rather than folding that job into the flagship.

The licence is more permissive than most, including Qwen's

The Kimi K3.1 Pro License opens with an MIT-shaped grant: use, copy, modify, merge, publish, distribute, sublicense and sell, plus run, deploy, fine-tune and create derivative works. Then it attaches two commercial conditions and exempts a third category.

ConditionWhat it requires
Model as a Service above $20MIf you or your affiliates run a Model as a Service business and aggregate revenue exceeds 20 million US dollars over any consecutive 12 months, you need a separate agreement with Moonshot before any commercial use
Attribution at scaleAbove 100 million monthly active users or $20M monthly revenue, "Kimi K3.1-Pro" must be displayed in the product interface
What Model as a Service excludesEnd-user products with model capabilities embedded in specific features or harnesses, and mere relaying of requests to models hosted by others
Full exemptionInternal use that exposes nothing to third parties, and any use through Moonshot's official products or certified inference partners

For the readers of this site, that is a comfortable document. Build a product on it, fine-tune it, sell what it generates, ship a coding tool, run it inside your company: nothing in the licence applies until you are running an inference business past twenty million dollars. Compare that with Qwen Community License 1.0, which reserves any independent commercial coding or office-productivity assistant with no revenue threshold at all, and K3.1 Pro is the more permissive of the two by a clear margin. It is a looser licence than MiniMax M3.1 Flash got as well.

It is still not MIT, and DeepSeek's competing model is. That gap is small for most projects and absolute for anyone who needs to redistribute without reading a licence.

What it does not do, starting with run on your hardware

You are not running this locally. 2.8T parameters at four bits per weight is on the order of 1.4 TB of tensors before any KV cache, and the 1M context is the feature most likely to need the memory you do not have. The realistic paths are the hosted API or a multi-node cluster with expert parallelism under vLLM, SGLang or TokenSpeed. There is no GGUF, no Ollama tag, and no sensible single-machine configuration, which is a change from what we were able to write about K3.

There is also no technical report, no training-data description, and no published form of xdev-bench-hard. Moonshot has not said whether fine-tuning can restore the removed capability, which was a fair question when only the API existed and is a live one now that the weights ship ungated under a near-MIT licence and the lab has itself measured the unrestricted ceiling at 90.5.

Two API-level gotchas worth knowing before you integrate. Thinking effort is a top-level reasoning_effort field taking low, high or max, and it defaults to max, so you are billed for the most expensive setting unless you say otherwise. And on multi-turn or tool-calling flows the complete assistant message must go back into messages unchanged, including reasoning_content and tool_calls, not just content. Strip those and the model loses its own reasoning thread.

Running it

export MOONSHOT_API_KEY=your_key_here

# the model is served on an Anthropic-compatible path as well as OpenAI's
curl -s https://api.moonshot.ai/v1/chat/completions \
  -H "Authorization: Bearer $MOONSHOT_API_KEY" \
  -H 'content-type: application/json' \
  -d '{"model":"kimi-k3.1-pro",
       "reasoning_effort":"low",
       "messages":[{"role":"user","content":"hi"}],
       "max_tokens":64}'

# read the licence before you plan a business on it: clause 2 is the only gate
curl -sL https://huggingface.co/moonshotai/Kimi-K3.1-Pro/raw/main/LICENSE

# check the weight format yourself, the card and the HF tags disagree
curl -sL https://huggingface.co/moonshotai/Kimi-K3.1-Pro/raw/main/config.json \
  | head -40

# size the download before you start it
curl -s https://huggingface.co/api/models/moonshotai/Kimi-K3.1-Pro \
  | grep -o '"rfilename":"[^"]*safetensors"' | wc -l

Set reasoning_effort explicitly on the first call you ever make. The default is max, and on a model priced per million tokens that is the difference between a cheap test and a surprising invoice.

Who should run this

If you are choosing an open agentic model for long-horizon software work, this is now the strongest option with a licence you can live with. SWE Marathon, Terminal-Bench 4.0 and HLE with tools are genuine wins over both Claude entries and GPT-6 Astra on Moonshot's sheet, and the licence stays out of your way below twenty million dollars of inference revenue. Benchmark it against DeepSeek V4.1-Flash on your own tasks, because that model wins DeepSWE and is MIT.

If you are writing about AI safety or tracking cyber capability, use the 33.3 figure. The 90.5 to 81.0 step is a measurement of a decision inside one lab. The 33.3 to 81.0 step is a measurement of what changed in the world, and it is the one with a licence attached and no access gate. Both come from Moonshot's own table.

If you were waiting for the weights to evaluate the cyber claims independently, you can start, but not on this benchmark. xdev-bench-hard is unpublished, so the only reproducible work is on ExploitBench and ExploitGym, where the published figures are cross-vendor self-reports under different harnesses and budgets. A single-harness run of K3.1 Pro, K3 and DeepSeek V4.1-Flash on one public cyber benchmark would be the most useful thing anyone could publish about this release, and nobody has.

Sources and further reading

The ten-minute version: send one request with reasoning_effort set to low, then the same request to DeepSeek V4.1-Flash, and compare on a task you actually have. The two models are 4.7 points apart on Moonshot's cyber benchmark and 1.5 points apart on DeepSWE in DeepSeek's favour, so for most workloads the deciding factor is not capability but whether you want a licence with a clause 2 in it.

Not independently tested. Every repository statistic, licence clause, architecture figure and benchmark score above was read from the public Hugging Face API and the raw model card, config and LICENSE files on the date below. All benchmark numbers are Moonshot's own, reported at maximum reasoning effort, with competitor figures taken by Moonshot from those vendors' published reports rather than measured by Moonshot; the evaluation footnotes confirm the harnesses and time budgets differ between columns. The xdev-bench-hard benchmark is internal to Moonshot and unpublished, so none of it is externally reproducible. Licence summaries are our reading of the operative text and are not legal advice.
Date checked: 2026-11-16

Prev Article
Edge0-145B
Next Article
OpenThinker-32B

Related to this topic: