TL;DR
Kimi has introduced K3.1 Pro, a post-trained model based on K3 with a 1-million-token context window, alongside K3.1 Pro Highspeed, which the company says runs 2–3 times faster. Both are available through Kimi’s products and API; the company says full model weights will be released by Nov. 4. Its benchmark results and performance claims come from Kimi, including results on internal tests.
Kimi has launched K3.1 Pro, a new model with a 1-million-token context window, and K3.1 Pro Highspeed, which the company says delivers two to three times the speed of the standard version. The models are available now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API; Kimi says it plans to release the full model weights by Nov. 4.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Kimi describes K3.1 Pro as the first model in its K3.1 series and a post-trained version of K3, developed with substantially more post-training compute. The company says it adjusted its reinforcement-learning framework, long-context inference, training efficiency and other systems to support training across hundreds of downstream domains. These are Kimi’s descriptions of its development process; the supplied announcement does not provide an independent evaluation of them.
The company reports a 65.7% score on Terminal-Bench 4.0, a test of coding agents on terminal tasks. In Kimi’s comparison, that is above the listed 58.2% score for GPT-6 Astra and just below Claude Opus 5.5 at 66.4%. On Kimi Code Bench 3.0 Long-horizon, which Kimi says is an internal test, K3.1 Pro scored 61.2%, compared with 55.95% for K3.
Kimi also highlights results in 3D modeling and finance. It reports a 92.7% score on BenchCAD with tools, compared with 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra. On OfficeQA Pro v2, which tests analysis of financial documents, Kimi reports a score of 64.4. The source does not specify enough about every benchmark’s scoring scale or evaluation conditions to establish how these results translate to typical use.
Faster Access to Long-Context AI
The release gives users and developers access to two versions of a model designed for tasks that can involve long inputs, coding tools, 3D workflows and financial documents. The one-million-token context window is intended to let the model process a large amount of material in a single context, while the Highspeed option offers a faster alternative according to Kimi.
For developers, the announcement also sets a near-term milestone: Kimi says the full weights are due by Nov. 4. If released as stated, they could allow broader access to the model beyond Kimi’s hosted services, subject to the eventual license and technical requirements. Those details are not included in the source material.
The performance figures may help readers compare the model with other systems, but they are company-reported results rather than independent evidence of performance across all tasks. Kimi identifies some tests as internal, and benchmark scores alone do not show reliability, cost, speed under particular workloads or suitability for high-stakes financial decisions.
Products Worth Considering
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
Bose QuietComfort Wireless Bluetooth Headphones – Active Over‑Ear Noise Cancelling with Mic, USB‑C Charging, Deep Bass, up to 24 h Playtime,
Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, Silver
As an affiliate, we earn on qualifying purchases.
K3.1 Builds on the K3 Model
Kimi presents K3.1 Pro as an update to its existing K3 model, not as a separate product line unrelated to that system. The company attributes the claimed gains to additional post-training compute and large-scale reinforcement learning, alongside changes to infrastructure used for training and evaluation.
The announcement describes examples of extended coding work, including an internal test in which the model iteratively developed a simulated walking controller. Kimi says the work involved 167 missions and 420 independent reviews, and that the simulated figure eventually walked for 11.2 seconds. It also says the model identified and corrected 36 errors in shoulder-muscle calculations after reward adjustments did not improve arm motion. These figures describe a company-reported demonstration, not a general measure of coding performance.
Kimi also says its 3D Make feature is live on kimi.ai and its app, allowing users to preview and edit generated models and export them for uses including 3D printing. The announcement links that feature to K3.1 Pro’s design capabilities, but does not detail its availability by region, subscription tier or device.
“Kimi K3.1 Pro is the first model in our new K3.1 series.”
— Kimi
Weight Release and Testing Details
The announcement does not specify the year for the Nov. 4 deadline, the weights’ license, or whether release will include all components needed to run the model. It also gives no independent confirmation that the weights will be published on schedule.
Kimi’s benchmark claims need to be read with their stated scope. Some results come from internal benchmarks, and the supplied material does not provide full test protocols, sample sizes, compute settings or independent replication. The company also makes a general claim of two-to-three-times higher speed for Highspeed without specifying the baseline conditions, latency measures or tasks tested.
The source text ends before completing a claim about factual errors on Kimi Hallu Bench Hard. It gives a claim-level hallucination-rate comparison of 23.4% for K3 and 8.3% for K3.1 Pro, but the following statement about the share of responses containing at least one factual error is incomplete. That missing figure cannot be reported from the supplied material.
Watch for the Weight Publication
The next stated milestone is Kimi’s planned release of the full model weights by Nov. 4. Readers and developers will be able to assess the release more fully once Kimi publishes the files, license terms and instructions for using them. The source does not identify a separate date for those details.
In the meantime, Kimi says both versions can be used through its web, work, coding and API services. Further information that would clarify the launch includes pricing, usage limits, measured Highspeed latency and independently reproducible benchmark methods. No additional release schedule or third-party assessment is included in the material provided.
Key Questions
What are Kimi K3.1 Pro and K3.1 Pro Highspeed?
K3.1 Pro is a post-trained version of Kimi K3 with a 1-million-token context window, according to Kimi. K3.1 Pro Highspeed is a faster version that the company says delivers two to three times the speed of the standard model.
Where can users access the models?
Kimi says both models are available now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API. The announcement does not give pricing or usage limits.
When will Kimi release the model weights?
Kimi says it will release the full weights by Nov. 4. The supplied announcement does not specify the year, license terms or publication platform.
Are the benchmark results independently verified?
The supplied figures are reported by Kimi. The company identifies some evaluations as internal, and the material does not provide independent replication or complete testing protocols.
What does the 1-million-token context window mean?
It is the context capacity Kimi advertises for K3.1 Pro, indicating the amount of text or other supported input the model is designed to handle within a session. The announcement does not explain practical limits, such as performance or cost at the maximum context length.
Source: hn
