Kimi K3.1 Pro is available, but open weights must wait

What this is about

Moonshot AI is offering Kimi K3.1 Pro as another option for developers who want capable models and, over time, less dependence on a single cloud provider. The crucial distinction is availability: the service can already be accessed, but the full model weights cannot. Moonshot promises to release them by November 4, 2026. Describing the model as already freely downloadable gets ahead of that promised release.

Developer Simon Willison documented API access on October 23, 2026, including a hands-on experiment. The official announcement page carries an October 24 modification date. This article reflects the situation on October 25, 2026. The news therefore concerns an operational API and a pending self-hosting promise, not two options that are already available together.

What Kimi K3.1 Pro actually does

According to the official announcement, the model builds on Kimi K3 with additional post-training. Its intended uses include programming tasks, creating 3D models, and analyzing financial documents. The advertised context window holds around one million tokens. Tokens are chunks of text: a large window provides room for extensive inputs, but does not automatically mean every detail will be understood correctly.

The model is accessible through Kimi products and the API. Moonshot also offers a Highspeed version that it says responds two to three times faster. That is a vendor claim, not a speed measurement performed by Cyber Ivy. For the standard version, Willison and Artificial Analysis list prices of $2 per million input tokens and $10 per million output tokens. Discounts for cached inputs and other billing details can change the final bill.

Why it matters

When checked on October 25, Artificial Analysis listed a score of 55 on its Intelligence Index. This provides an independent reference alongside the vendor's benchmarks. It does not establish that the model outperforms competitors on every task. The more interesting combination is measurable capability, API access that can be tested now, and a promised opportunity to run the weights independently later.

For a small development team, another usable model could increase flexibility around costs, data processing, and technical dependence. That benefit only materializes if the model reliably solves the team's actual tasks. A low token price is of little help if repeated failures and manual corrections make the work more expensive. Cost per accepted solution is consequently more informative than an isolated list price.

The promised weights are also not equivalent to an unrestricted open-source release. What is actually permitted and economically practical must be checked against the published files, license terms, and operating requirements. Until then, self-hosting remains a prospect rather than an alternative available today.

In plain language

Imagine a well-equipped kitchen available for rent. Today you can cook there and pay for using it. The operator promises to provide the blueprints and equipment list later. That could make it possible to build your own kitchen, but you would still need premises, appliances, and appropriate usage rights. Access to the rental kitchen does not establish that you can already build your own.

In the same way, Kimi's immediate internet-based access is separate from running downloaded model weights later. Both approaches can be useful, but they answer different questions about control and effort.

A practical example

A fictional development team tests 100 clearly defined programming tasks. Before starting, it decides that a solution only counts if it passes existing tests and human review. In this illustrative calculation, each task uses 10,000 uncached input tokens and 2,000 output tokens. That adds up to one million input tokens and 200,000 output tokens. At the quoted prices, the total is $2 plus $2, or $4 in token charges alone.

This calculation is not a prediction for real agent runs. Additional reasoning and output tokens, retries, tool calls, taxes, and labor are excluded. The team therefore also records waiting time, rework, and accepted solutions. Only then does it compare Kimi with its existing model. It uses non-sensitive test code for the experiment and does not grant write access to production systems.

Scope and limits

• The weights are still pending. Moonshot names November 4 as its promised deadline. That does not establish whether the eventual release, license, and operating effort will meet a particular user's requirements.
• Benchmarks do not replace acceptance testing. An index score measures a particular set of tasks. It guarantees neither correct financial analysis nor secure code changes. Expert review and task-specific testing remain necessary.
• API access does not provide local data privacy. Confidential data should only be submitted after reviewing the relevant contractual and processing terms. Even available weights would not guarantee free infrastructure or easy operation on a laptop.

The hero photograph is illustrative, not a picture of Moonshot's infrastructure: Carl Lender, “Datacenter Server Racks,” via Wikimedia Commons, CC BY 2.0. Reduced-size version, with no substantive alteration.

SEO & GEO keywords

Kimi K3.1 Pro, Moonshot AI, open model weights, AI API, coding agents, Artificial Analysis, model comparison, token pricing, self-hosting, AI development