The company claims a million-token context window for Pro and promises to release the model weights by November 4, 2026, enabling users to run it on their own infrastructure.
Moonshot says its superintelligence (SI) model Kimi K3.1 Pro and its Highspeed variant are available through its applications and developer interface. On 26 October, the company apologised for rate limits and unnecessary consumption of subscribers’ usage allowances after the launch and offered quota relief.
The 23 October launch built on the earlier Kimi K3 model through additional training, for which Moonshot says it used substantially more computing power.
Moonshot said it was offering all subscribers a 25% monthly-quota reset, which it described as equivalent to one week’s usage. Subscribers who had used less than 25% would have their recorded usage reset to zero. CNA carried the company’s announcement on 26 October.
The apology covered rate limiting—restrictions on how frequently users can make requests—and unnecessary quota consumption. Moonshot also said it had adjusted operating settings to reduce unnecessary subagent calls, in which an SI system delegates work to other agents, and to shorten generated responses. These are company-reported changes intended to reduce usage consumption, not independently verified improvements.
SiliconANGLE’s 24 October report said Pro’s cybersecurity benchmark scores significantly trailed GPT-6 Astra and Anthropic’s Claude Opus 5.5. Moonshot says it deliberately limited cybersecurity training to address safety risks. It says Pro’s cyber capability remains below that of leading proprietary models—whose weights are not publicly released—that had been publicly available for at least three months. These comparisons are reported evaluations, not independent verification of performance.
The W27 keeps the Galaxy Z Fold 8 Ultra’s core hardware, with dark-red or dark-blue finishes, golden accents and 16GB of working memory in both storage versions.
2 min
Moonshot says it removed cyber-related data from an intermediate training stage and from reinforcement learning, which trains a model using feedback on its attempts. It retained general-purpose coding tasks, with the stated aim of reducing misuse risk while preserving general usefulness.
The company compared Pro with a training version that lacked those controls on xdev-bench-hard, its internal test of finding and exploiting software vulnerabilities. Tasks supplied vulnerable code and a bug hint; the model had to exploit the flaw in an isolated test environment. The scores measure success within three attempts, using a random subset of runs without hard time limits or limits on output tokens—the units of material a model generates.
| Training version | Success within three attempts |
|---|---|
| Without the cyber-data controls | 90.5% |
| Final Kimi K3.1 Pro | 81.0% |
Moonshot says Pro’s cyber capability nevertheless improved substantially over K3 because stronger general coding and extended debugging skills carried over to cyber tasks. The reduction shown in its test is relative to the uncontrolled training version; it does not establish that exploit capability was eliminated or that the model is safe.
Moonshot says its release policy aims not to push open-weight models beyond the highest available cyber capability, while progressively increasing the capability available through open weights. SiliconANGLE also said the company plans more advanced vulnerability-analysis features in future models.
The Decoder’s 23 October report said Pro scored 55 points on the Artificial Analysis Intelligence Index, gaining 11 points over K3 and placing fourth. It also said Pro’s token prices fell by 33%, but that reduction did not translate into a similar saving per evaluation task.
Tokens measure both material sent to a model and the response it generates. Input prices cover the former; output prices cover the latter. Cached input is previously processed material that the service can reuse.
| Measure | Unit | Kimi K3 | Kimi K3.1 Pro |
|---|---|---|---|
| Input price | Per million tokens | $3.00 | $2.00 |
| Output price | Per million tokens | $15.00 | $10.00 |
| Cached-input price | Per million tokens | $0.30 | $0.20 |
| Intelligence Index evaluation cost | Per task | $2.00 | $1.98 |
The Decoder attributed the nearly unchanged task cost to Pro generating 14% more output tokens. For developers, the token rate is a price per unit, while a task’s total cost also depends on how many tokens the model uses.
Moonshot reports improvements in coding, 3D creation and financial research. On Kimi Code Bench 3.0 Long-horizon, its internal test of multi-step coding, tool use and end-to-end problem solving, Pro scored higher than K3. These figures are Moonshot’s own measurements.
| Model | Score |
|---|---|
| Kimi K3 | 55.95% |
| Kimi K3.1 Pro | 61.2% |
Moonshot also reports the following Terminal-Bench 4.0 scores. That test evaluates coding agents—models that use tools to carry out tasks—on command-line work with limited computing resources and task-specific checks.
| Model | Score |
|---|---|
| Kimi K3.1 Pro | 65.7% |
| GPT-6 Astra | 58.2% |
| Claude Opus 5.5 | 66.4% |
Moonshot reports that Pro scored 92.7% with tools on BenchCAD, which tests reconstruction of 3D objects from multiple-view images by generating computer-aided design (CAD) code. SiliconANGLE said this put Pro within four percentage points of GPT-6 Astra and Claude Opus 5.5. These are reported results, not independently verified performance.
Moonshot announced both variants on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, which lets developers connect the models to their software. It also announced Pro’s rollout on Kimi mobile apps. SiliconANGLE described launch access as being through Moonshot’s paid cloud services.
Moonshot says Pro has a one-million-token context window—the limit on how much material it can work with at once. The company says Highspeed delivers two to three times Pro’s speed.
Kimi Enterprise access to Pro was off by default at launch; administrators can enable it for their organisations. Individual users can select Pro in Kimi Code with the /model command, while developers can select either variant through the Kimi API.
Moonshot commits to releasing the full model weights—the learned numerical settings needed to run the model—by 4 November 2026. That release would enable users to run Pro on infrastructure they control.
The satellite operator expects access to more than twice OneWeb’s current bandwidth in return for spending in 2027–2034, with initial IRIS² services expected from 2030.
3 min
The black, gold and green edition follows September’s sold-out preorders. Nintendo’s replenished launch-day stock also quickly sold out.
1 min
