Microsoft has added Moonshot AI's Kimi K3.1 Pro to its Foundry model catalog, giving developers a managed way to run the 2.8-trillion-parameter open-weight model without downloading and serving its 1.56 terabytes of weights themselves. The listing, published on 5 November, covers the text-and-image model with a context window of up to one million tokens, priced at $2.30 per million input tokens, $0.23 cached, and $11.50 per million output tokens.
What lands on Foundry
Kimi K3.1 Pro is the first model in Moonshot's K3.1 series, a post-trained version of Kimi K3 built for agentic work. It is a mixture-of-experts system with 2.8 trillion parameters, of which 104 billion are active for any given token. A 401-million-parameter vision encoder, MoonViT-V2, lets it read images alongside text, and the context window stretches to 1,048,576 tokens.
Moonshot published the model's weights on Hugging Face on 4 November, keeping the date it set. The download runs to 96 files and about 1.56 terabytes, which is why a managed endpoint matters: nobody can host a model of this size at home. The weights ship under Moonshot's custom license, which asks large model-as-a-service operators to sign a separate agreement.
The cloud pattern
Foundry is not the model's first cloud home. Amazon added Kimi K3.1 Pro to Bedrock days after the weights landed, and Microsoft's post places it alongside open-weight models from NVIDIA, OpenAI, DeepSeek, MiniMax, Mistral AI, and Qwen. The catalog keeps growing: Microsoft added five video models from fal to Foundry in the same week.
The pricing is straightforward: $2.30 per million input tokens, $0.23 for cached input, and $11.50 per million output tokens. A request with 100,000 uncached input tokens and 10,000 output tokens costs about $0.35, or $0.14 if all the input qualifies for the cached rate. For teams already running workloads on Foundry, the model slots into billing, identity, and monitoring controls.
The benchmarks, with a caveat
Moonshot's numbers put Kimi K3.1 Pro within a point of Claude Opus 5.5 on Terminal-Bench 4.0, scoring 65.7 against 66.4, and ahead of GPT-6 Astra at 58.2. On SWE Marathon v1.1, a long-horizon software engineering test, it leads the field at 51.7. Independent rankings agree the model is near the top: it sits fourth on the Artificial Analysis Intelligence Index.
The comparisons come with a warning, and Microsoft states it: figures are a starting point, not an answer. Moonshot ran its model on one harness while taking rival scores from reports and leaderboards, so the numbers are not fully controlled. Microsoft's advice is to test the model on tasks and judge the quality of the work, not the first reply.
Our read
The significance here is distribution, not the model itself. Kimi K3.1 Pro was a strong release, but a 1.56-terabyte download kept it out of reach for most teams. Foundry turns it into an API call inside a cloud many enterprises already use, with procurement, compliance, and billing in place. That converts an open-weight release into a model teams can evaluate.
There are limits. The license bites for large operators, and Moonshot trimmed the model's cyber capabilities before release, cutting its score on Moonshot's internal security test from 90.5 to 81.0. A Chinese lab's flagship on a US cloud invites policy questions, though Microsoft frames the listing as its open-weight investment. The test is adoption: whether teams pick it over the models they run.
Why It Matters
The change is choice without the plumbing. Foundry lets a team test Kimi K3.1 Pro against the models it uses, on the same billing and governance, and keep it only if it finishes work better. For Moonshot, the listing is enterprise reach it could not buy alone: the model sits where customers shop, next to its hosted-agents push for users.
Background
Kimi K3.1 Pro launched on 23 October in Moonshot's apps and API, with the full weights promised by 4 November. Moonshot kept that date, publishing the 1.56-terabyte checkpoint on Hugging Face on schedule. Microsoft's Foundry listing followed a day later, on 5 November, completing a two-week run from launch to open weights to a managed enterprise endpoint.
Source: Microsoft's Foundry announcement and the Kimi K3.1 Pro model card
0 Comments