This Kimi K3.1 Pro review has to start by separating three claims that headlines tend to merge. A one-million-token context window is about what the model can read. Open weights are about what you may download. Neither says anything about what you can run, and for this model the gap is wide: the published checkpoint is reported at roughly 1.56 TB.

Moonshot AI introduced the model on October 23, 2026 through its apps and API, and published the weights on Hugging Face on November 4, as it had promised. Since then, the useful question is no longer whether the model is good. It is who benefits from a download this large, and how far the company's own numbers can be trusted. We read Moonshot's announcement, checked its claims against independent leaderboards, and scored the model with the ToolRadar weighted formula, with no claim of hands-on testing. Our earlier coverage of the family is in our Kimi K2.6 review.

Mahmoud Salamoun

Founder, ToolRadar · Reviewed Nov 5, 2026

Independent AI tools reviewer with a background in marketing and content, and hands-on daily experience directing AI tools like Gemini and ChatGPT for real work. This review is based on official documentation, published pricing, and verified third-party coverage — not a claim of hands-on testing of Kimi K3.1 Pro itself unless stated otherwise.

What Kimi K3.1 Pro Is

K3.1 Pro is a post-trained upgrade of Kimi K3, which Moonshot introduced in July. Moonshot credits the gains to more post-training compute and a reworked reinforcement-learning stack, according to VKTR's coverage. It accepts text and images and replies in text, and it is aimed at coding and multi-step agent work.

Two versions exist. The standard model is kimi-k3.1-pro in the API. A Highspeed variant, kimi-k3.1-pro-highspeed, is marketed at two to three times the speed, a figure that comes from Moonshot. You can reach the model through Kimi's web and mobile apps, the Kimi Work desktop app, Kimi Code, and the API. On Kimi Enterprise, administrators must switch it on, because it is off by default.

Open Weights and the License

According to OrcaRouter, the repository is moonshotai/Kimi-K3.1-Pro, ungated, with 96 safetensors shards totaling about 1,561 GB. The same report says Moonshot never shipped a BF16 checkpoint: the weights are MXFP4 with MXFP8 activations, produced through quantization-aware training. That matters because earlier coverage estimated 5.6 TB at BF16, a number that turns out to describe a file that does not exist.

"Open-weight" is the right word, and "open-source" is not. The license is custom and not approved by the Open Source Initiative, per OrcaRouter, with the repository metadata reading license: other. The terms it summarizes are these:

• Model-as-a-service: if a company gives third parties inference or fine-tuning access and its group revenue exceeds $20 million over any 12 consecutive months, it needs a separate agreement with Moonshot first. Embedding the model in a product feature is carved out.
• Attribution: commercial products above 100 million monthly active users, or $20 million in monthly revenue, must display "Kimi K3.1-Pro" prominently.
• Exemptions: internal use that does not expose the model or its outputs to third parties, and use through Moonshot's official products and certified inference partners.

For a startup or a developer, those thresholds are far away. For a company that sells hosted access to models, they are not. Treat this as a summary and read the license text yourself. The previous generation set the pattern: Simon Willison noted in July that the K3 license "no longer calls itself 'modified MIT'" and added a separate-agreement requirement for large service businesses, as he wrote in his post on K3. Whether K3.1 Pro's license is word-for-word the same, we could not confirm.

The Benchmarks

This table is Moonshot's own, relayed by VKTR, so treat it as a claim and not a result. Moonshot states it ran at maximum reasoning effort, temperature 1.0 and top-p 0.95, and that some rival scores came from third-party leaderboards or technical reports.

The pattern is a split verdict. Kimi is within 0.7 points of Opus 5.5 on Terminal-Bench 4.0 and ahead on two rows, and it is well behind on Program Bench and the science variant. The rows that favor it are the ones a developer might care about most, which is not a reason to ignore the rows that do not.

Independent numbers are thinner but point the same way. On the Vals AI leaderboard, K3.1 Pro scored 62.52% on the Vals Index, ranked 9th of 32 on the page we read, and 63.64% on Terminal-Bench 4.0, ranked 3rd. It was far lower on others: 37.5% on Legal Research Bench (43rd of 89) and 24.87% on Tax Agent Bench (42nd of 80). Latent Space's October 24 roundup reports Artificial Analysis placing it fourth on its Intelligence Index at 55, at about $1.98 per index task. Those index numbers change with the index version, and a later article cites a different, lower score on a newer version, so do not compare figures across versions.

One more caution from the same roundup: Vals found that for the earlier K3 model, accuracy on a legal test varied by provider, 60.6% through Baseten versus 75.8% through Moonshot. If you route K3.1 Pro through a third-party host, test that host, not the model in the abstract.

Running It Yourself

This is where the headline and the reality separate. Per OrcaRouter's summary of the vLLM recipe, self-hosting needs vLLM 0.34.0 or newer and a single group of eight GPUs, described as one B300 node or two GB300 trays. An eight-GPU H200 node, with about 1.13 TB of memory in total, falls short of the 1.56 TB of weights before any KV cache is counted. Weight size is only part of the memory bill, because the context cache grows with length and concurrent users.

The same source estimates a suitable node at roughly $550,000 to $650,000 and compares that to API list prices. It sells API routing, so its conclusion favors the API, and the estimate leaves out power, staff, and idle time. We cannot verify it, but it is directionally consistent with a model of this size. The practical reading is simple: no consumer machine runs this, so tools like Ollama are not the route. If you want to rent GPUs instead, our notes on RunPod's real cost and Modal's serverless GPUs show how rented capacity gets priced, and our guide to running AI agents 100% locally covers models that do fit at home.

What the open weights do buy is control. An organization with the hardware can audit the model, fine-tune it, and keep data inside its own walls. Everyone else gets the main benefit indirectly: competition, and a price floor for hosted access.

Pricing and Highspeed

API prices below match across the two summaries we read, both dated October 24 to 25, which cite Moonshot's price list. Verify them on the live page before building a budget.

Highspeed charges double for a claimed two to three times the speed. That trade only makes sense where waiting costs you something, such as an interactive coding assistant. One summary says the Batch API guide did not list K3.1 Pro, so do not assume a batch discount. Kimi's subscription tiers exist, but the sources we read disagree on their prices and some quote annual rates as monthly, so check the live page. Subscriptions do not cover API usage, which is billed separately.

Cyber Limits and Privacy

Moonshot says it found cyber-related material in its mid-training data and its reinforcement-learning environments and removed it for the open release. On its internal xdev-bench-hard exploit suite, an unrestricted checkpoint solved 90.5%, the released model 81.0%, and the older K3 33.3%, according to Moonshot. Reuters covered the decision in an October 24 report. These are internal numbers with no independent reproduction. And a 9.5-point drop on a suite where the model still far exceeds its predecessor is a reduction, not a removal. Latent Space notes Artificial Analysis gave K3.1 Pro a 64% on its Cyber Index, second overall behind a trusted-access model at 69%.

On data, Moonshot offers Zero Data Retention for API customers. We could not verify the terms for any specific account. An API call sends your prompts to the provider, while self-hosted weights keep them in your own environment, and the Kimi apps should not be assumed to follow the API's retention rules. Moonshot also states that its hallucination benchmark, which showed claim-level errors falling from 23.4% to 8.3%, is a stress test that "does not represent" everyday error rates, and VKTR notes GPT-6 Astra scored 5.5% on the same test.

Pros and Cons

💡 What Reviewers Say

The first quote is about the earlier K3 license, and the second is a smoke test, not a benchmark, so neither settles anything. Together they show the mood: respect for the model, caution about the terms. The same Latent Space roundup records other reactions. Nathan Lambert noted the model scoring above some rivals on Artificial Analysis' index, and a poster known as giffmana warned against calling trillion-parameter models "local" because they need serious accelerators. Another poster reported a game-based test where performance resembled DeepSeek V4.1 Flash at a cost comparable to Argon, which Latent Space presents as one anecdote. We found no verified Hacker News thread with substantive comments to quote, so we cite none.

Who It Fits

Expert Editorial Opinion

Final Verdict

Final score = (8.3 × 0.30) + (8.3 × 0.25) + (7.4 × 0.20) + (7.9 × 0.15) + (7.5 × 0.10) = 2.49 + 2.075 + 1.48 + 1.185 + 0.75 = 7.98, shown as calculated and not rounded up.

❓ Frequently Asked Questions

🔗 Related ToolRadar Reviews