AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: The Kimi K3.1 Pro Result Is Close To The Frontier, But The Weights Aren’t Here on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get audio and creator gear delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

Moonshot AI launched Kimi K3.1 Pro on October 23, and Artificial Analysis ranked it fourth on its Intelligence Index, close to leading proprietary models. The model is not yet available as downloadable weights: Moonshot says those are due by November 4, and has not published a K3.1 licence. Independent tests support some performance claims, while several headline capabilities and a 65% hallucination reduction come from Moonshot’s own evaluations.

Moonshot AI launched Kimi K3.1 Pro on October 23, and Artificial Analysis ranked it fourth on its Intelligence Index, three points behind the leading model in the comparison. The result puts Kimi close to top proprietary systems, but it is not yet an open-weights release: Moonshot says downloadable weights are due by November 4, and has not published the model’s licence.

Artificial Analysis scored K3.1 Pro at 55.1, compared with 57.6 for Claude Opus 5.5, 56.0 for Claude Sonnet 5.5 and 55.8 for Muse Wonder Full. The evaluator’s model page lists K3.1 Pro with 2.8 trillion total parameters and 104 billion active parameters, using a mixture-of-experts architecture. It supports text and image input and has a listed one-million-token context window.

The independent results show especially strong performance on agentic tasks. Artificial Analysis reported a 67% score on Terminal-Bench 4.0, compared with 13% for the earlier Kimi K3. On GDPval-AA, a simulated knowledge-work evaluation, K3.1 Pro scored 1,846 Elo, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. Its score on Humanity’s Last Exam was 52%, while Opus 5.5 scored 61%.

Artificial Analysis lists API pricing at $2 per million input tokens and $10 per million output tokens. It estimates K3.1 Pro costs $2.11 per Intelligence Index task, compared with $5.98 for Opus 5.5 at its highest reasoning setting. Those figures reflect the evaluator’s stated comparison and do not establish that every real-world workload will cost one-third as much.

At a glance
reportWhen: Launched October 23, 2026; weights prom…
The developmentKimi K3.1 Pro has received a strong independent benchmark ranking, but its promised open weights and licence are not yet available.
Thorsten Meyer AI
Moonshot AI · 23 Oct 2026

Infographic 01 · Claimed vs. measured

Moonshot says “matches the strongest.” Independent tests say: close.

The KPIs at the top are independent measurements. Vendor numbers and independent numbers come from different harnesses, so they are shown side by side, never mixed.

55Intelligence Index score: 4th of all models
3 ptsbehind the leader, Claude Opus 5.5 (58)
67%Terminal-Bench 4.0: highest AA has measured
$2.11per index task, vs. $5.98 for Opus 5.5 (max)

Moonshot reports

vendor claims own harness, partly internal tests

BenchmarkK3.1 ProComparison
Terminal-Bench 4.065.7%vs. Claude Opus 5.5: 66.4%
BenchCAD with tools92.7%vs. Claude Opus 5.5: 96.2%
OfficeQA Pro v264.4vs. Claude Opus 5.5: 62.2
Kimi Hallu Bench Hard internal8.3%vs. Kimi K3: 23.4% (about −65%)

Artificial Analysis measured

independent highest reasoning setting per model

MeasurementK3.1 ProComparison
Terminal-Bench 4.067%vs. Claude Opus 5.5 (max): 60%
GDPval-AA1,846 Elovs. Claude Opus 5.5: 1,866
Humanity’s Last Exam52%vs. Claude Opus 5.5: 61%
AA-Omniscience hallucination rate44%vs. Kimi K3: 53% · Gemini 4 Argon: 15%

The 65% hallucination drop holds on Moonshot’s own stress test. On an independent knowledge test, the rate fell from 53% to 44%.Moonshot: the stress test rates “do not represent overall error rates in daily use.”

Sources: Moonshot AI, Kimi K3.1 Pro tech blog (23 Oct 2026) (left card, vendor charts); Artificial Analysis launch article (23 Oct 2026) and Artificial Analysis model page (data as of 24 Oct 2026) (KPIs and right card).
Powered by Thorsten Meyer AI

A Near-Frontier Model, Not Yet Open

The result matters because a model with downloadable weights can give organisations more control over where and how they run AI, including the option to keep data within their own infrastructure. If K3.1 Pro’s weights arrive and its licence permits the uses companies need, its benchmark position could make it a notable alternative to paying for access to closed models.

For now, that possibility remains conditional. Benchmark performance is not the same as practical availability: users cannot yet inspect or run the promised weights, and the missing licence leaves questions about commercial use, modification and redistribution unanswered. The model’s scale also presents a substantial deployment barrier. At 2.8 trillion total parameters, running it independently is a major infrastructure undertaking, not a routine download for most individuals or small teams.

The cost comparison may interest developers choosing hosted APIs, while the large gain on Terminal-Bench suggests potential strength in tool-using coding work. But evaluation scores do not establish reliability across all software projects, nor do they show how the model performs under every company’s security, latency and workload requirements.

Products Worth Considering

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight

As an affiliate, we earn on qualifying purchases.

How K3.1 Compares With K3

K3.1 Pro is described in the source material as a more heavily post-trained version of Kimi K3. The earlier model scored 43.6 on the Artificial Analysis Intelligence Index; K3.1 Pro’s 55.1 is a substantial increase on that evaluation. Before the launch, the highest-ranked open-weights model in the cited comparison was Xiaomi’s MiMo-V2.6-Pro at 46.3, while Opus 5.5 scored 57.6.

Artificial Analysis calls K3.1 Pro a model that “will be the leading open-weights model once its weights are published.” The wording is conditional: the index result is available now, but the model cannot yet be treated as downloadable open weights. Artificial Analysis’ comparison uses each listed model at its highest reasoning setting, so the scores reflect those test configurations rather than every setting or deployment.

Moonshot also promoted examples involving game development in Unity, Unreal Engine and Godot, parametric modelling in Blender and CAD, and finance tasks. Those examples and several related results come from the company’s own demonstrations or internal benchmarks; they should not be conflated with the independent Artificial Analysis scores.

“The company describes its hallucination test as “deliberately crafted” to provoke errors and says the rates “do not represent overall error rates in daily use.””

— Moonshot AI

Weights, Licence and Real-World Results

K3.1’s weights are not yet available, and the promised November 4 date is a company commitment rather than a completed release. Moonshot had also not published a licence for K3.1 in the source material, so the terms for commercial use, modification and redistribution remain unknown. The absence of a licence makes it difficult for organisations to assess whether the eventual release will meet their needs.

Some prominent capability claims remain vendor-reported. Moonshot says the model completed projects across several game engines and improved a simulated walking task over 167 missions. Those examples were not independently verified in the cited material. Moonshot also reported a 65% reduction in claim-level hallucinations on its internal Kimi Hallu Bench Hard test, from 23.4% for K3 to 8.3% for K3.1 Pro. Artificial Analysis measured a smaller decline on its separate AA-Omniscience evaluation, from 53% to 44%. The tests use different methods, and neither result by itself establishes an overall error rate in ordinary use.

It is also not clear what hardware and operating costs would be required for organisations seeking to run the full model themselves, or how its benchmark performance will hold up on independent, real-world deployments. The stated scale suggests considerable infrastructure demands, but the source material does not provide a specific hardware configuration or cost estimate.

The November 4 Release Test

The next milestone is whether Moonshot releases the weights by November 4, as promised. A release would allow researchers and organisations to inspect the files, test deployment requirements and examine the published licence. Until then, K3.1 Pro is available through the Kimi apps, Kimi Work, Kimi Code and the Kimi API, rather than as a downloadable model.

Readers should watch for the actual licence and for independent evaluations after the weights become available. Those would help establish whether the benchmark result can be reproduced outside the company’s hosted services, what kinds of use are permitted, and how much infrastructure the model requires. If Moonshot misses the announced date or publishes restrictive terms, the practical significance of the open-weights claim will change.

Key Questions

What did Artificial Analysis find about Kimi K3.1 Pro?

It scored 55.1 on the Intelligence Index and ranked fourth in the comparison cited in the source material, behind Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full. Results varied by test, with a particularly high 67% on Terminal-Bench 4.0.

Can users download Kimi K3.1 Pro now?

No. The source material says it was available through Moonshot’s apps and services, while the weights were promised by November 4. As of the report’s October 24 date, they had not been released.

Has Moonshot published a licence for K3.1?

No K3.1 licence had been published in the source material. Without it, commercial-use and redistribution terms remain unclear.

Is Moonshot’s 65% hallucination reduction independently confirmed?

No. That figure comes from Moonshot’s internal Kimi Hallu Bench Hard test. Artificial Analysis reported a smaller decline on its separate AA-Omniscience evaluation, from 53% to 44%; the results come from different tests and should not be treated as directly interchangeable.

Source: ThorstenMeyerAI.com

You May Also Like

Investigating The Generalization Of LLM-Engineered Agent Harnesses: ByteDance Seed’s Perspective

ByteDance Seed’s HarnessDev project tests if large language models can autonomously improve agent harnesses; only 34 of 64 modifications generalized beyond training conditions.

How To Secure AI Data: 2026 Checklist

A 2026 checklist from ThorstenMeyerAI.com outlines storage, encryption, power and labeling options for protecting AI project data and testing recovery.

The Google I/O 2026 Preview: What May 19-20 Will Reveal About Google’s Agentic Bet

Preview of Google I/O 2026 reveals expected launches of Gemini 4.0, multi-agent protocols, and XR glasses, shaping AI’s consumer and enterprise future.

The Cost Of Thinking Harder: Claude Fable 5.2 per-Task Pricing

Artificial Analysis reports Fable 5.2 scores from 54 to 65 as average task costs range from $3.10 to $26.54 across effort levels.