1300 166 138 Tell us what you need
News / Local AI

China’s Kimi K3.1 Pro nearly matches Claude Opus 5.5

By Monika Tantau • Published 3 weeks ago
A woman on a phone call types on a laptop at a home office desk, the kind of everyday office and coding work Kimi K3.1 Pro is pitched at

Key takeaways

• Moonshot AI says Kimi K3.1 Pro scored 65.7% on Terminal Bench 4.0, against 66.4% for Claude Opus 5.5
• It reads up to 1 million tokens, and a Highspeed version is said to run two to three times faster
• The full weights are now out, about 1.56 terabytes, so businesses can run it on servers they control
• Every score is Moonshot's own, and it admits the model trails top proprietary models on cyber tasks

Most Read

Moonshot AI, the Beijing company behind the Kimi chatbot, has released Kimi K3.1 Pro, and on the coding test it puts first it lands within a point of Anthropic’s Claude Opus 5.5. The model is already available in the Kimi apps and API, and the full weights are now out on Hugging Face, as Moonshot promised.

That second detail is what makes it worth an Australian business’s attention. Open weights mean a company can download the model and run it on servers it controls, instead of sending data to China. Every score below comes from Moonshot’s own announcement, and none has been checked independently yet.

What Moonshot’s new model can do

K3.1 Pro is a retrained version of Kimi K3, built with much more post training. It can read up to 1 million tokens in one go, roughly the length of several long novels, and it comes in two versions: the standard model and K3.1 Pro Highspeed, which Moonshot says runs two to three times faster.

Moonshot pitches it for coding, 3D and CAD design, finance and office work. It is available in the Kimi apps on iOS and Android, in the Kimi Work desktop app for Windows and Apple silicon Macs, in Kimi Code and through the API. On company accounts it starts switched off, and an administrator has to turn it on.

How close it gets to Claude and GPT 6

On Terminal Bench 4.0, which measures how well a model completes jobs in a computer terminal, K3.1 Pro scored 65.7%. Moonshot puts Claude Opus 5.5 at 66.4% and GPT 6 Astra at 58.2%. On BenchCAD, a 3D design test, it scored 92.7 with tools, behind Opus 5.5 on 96.2 and GPT 6 Astra on 95.9, and just ahead of Claude Fable 5.1 on 92.6.

It also makes fewer things up than its predecessor. On Moonshot’s hallucination stress test, the share of false claims fell from 23.4% in K3 to 8.3%. GPT 6 Astra still did better, at 5.5%, while Opus 5.5 scored 15%. Moonshot stresses that these prompts are built to be hard and are not error rates in everyday use.

A gloved hand holds a computer chip printed with the Chinese flag, a nod to the Chinese model whose maker admits it still trails on cyber tasks

What Moonshot admits it cannot do yet

The announcement is unusually frank about one weakness. Moonshot removed cybersecurity material from parts of the training and says K3.1 Pro’s cyber skills remain behind leading proprietary models. A test version trained without that removal scored 90.5 on its cyber benchmark, against 81 for the model it released.

The announcement did not include prices. The licence is now public, and it is Moonshot’s own rather than a standard open source one. Commercial use is allowed, with two conditions. A company selling the model as a service needs a separate deal with Moonshot once its revenue passes about $29 million in 12 months. And a commercial product with more than 100 million monthly users, or about $29 million in monthly revenue, must show the Kimi K3.1 Pro name prominently. Neither applies to internal use.

Why Australian businesses should care

A Chinese model raises an obvious question for local companies. The Australian Government banned DeepSeek from its own devices in February 2025 over security concerns, and many businesses apply similar caution to Chinese apps. Using K3.1 Pro through Kimi’s apps or API means sending prompts to Moonshot, even with the Zero Data Retention option it offers API customers.

The weights change that. An Australian team can now run the model on its own hardware or in an Australian cloud region, and no data has to leave the business. The download alone is about 1.56 terabytes, so this means serious server hardware, not a laptop, and our guide to the best local LLMs covers what that involves.

Where prompts are processed matters under the Privacy Act, as our explainer on AI data residency sets out. Matching the right model to each job, and deciding where it should run, is usually the first question in our AI consulting work.

Open models keep closing the gap

K3.1 Pro follows Meta’s Muse Spark Open and DeepSeek’s V4.1 Flash as another downloadable model that gets close to the best paid ones on selected tests. The pattern matters more than any single score: for a growing number of tasks, the choice is no longer between a top model and a weak one.

The sensible next step is to read the licence and test the model on your own work, ideally on a rented server before buying hardware. Benchmarks chosen by the company that built a model are a starting point, not a verdict.

Scroll to Top