Kimi K3.1 Release Signals Emerge as Moonshot Prepares Its Next AI Upgrade

A New Model Identifier Points to Moonshot's Next Move

Moonshot AI appears to be preparing Kimi K3.1 after developers found a callable model identifier, "kimi-k3-1", in the company's API platform registry on September 28. A preview subsequently appeared on the Kimi open platform under the names K3.1 or K3 11111, advertising a context window of one million tokens.

The discoveries suggest that development has moved beyond an internal name. Developers said the identifier responded to interface probes, while changes to Moonshot's website included another identifier, "k3d1-agent", and a new bot module. K3.1 is reportedly undergoing limited testing, sometimes called gray testing, in which access reaches a restricted group before a wider launch.

Moonshot's existing Kimi K3 has 2.8 trillion parameters, the numerical values a model learns during training. That figure belongs to K3; a parameter count for K3.1 has not been confirmed. The distinction matters because an upgrade can change how a model reasons, uses tools or handles requests without necessarily increasing its size.

The reported plans point toward longer autonomous assignments, coding, games, 3D development and computer operation. Possible additions include several reasoning settings and modes for coordinating multiple AI agents. Those features would push Kimi further toward completing sequences of work, rather than simply answering individual questions.

No firm public release date or final specification has been given for K3.1. For users deciding whether to adopt it, the practical questions are already taking shape: how reliably it completes work, how long it takes and how much that work costs.

What the September Discoveries Reveal

Developers first noticed "kimi-k3-1" in the API platform's backend model registry on September 28. An API is the connection that lets software send requests to an AI model and receive its answers. A model identifier tells that connection which system to use.

The subsequent platform preview added a more concrete capability claim: support for up to one million tokens of context. Tokens are small units of information, often words or parts of words. A context window is the amount of material a model can consider in one interaction, including instructions, documents and the conversation around a task.

The website identifier "k3d1-agent" and the newly added bot module have also prompted speculation about a Kimi Bot. Moonshot has not set out a confirmed product description for that possibility, so the clearest evidence concerns model testing and changes to the platform.

The sequence places the latest discoveries after K3's July introduction:

• July 17, 2026: Moonshot announced Kimi K3, with 2.8 trillion parameters and a context window of one million tokens.
• July 27: Moonshot's announced target for releasing K3's model weights, the learned numerical values needed to run the model.
• September 28: Developers found the callable "kimi-k3-1" identifier in the API registry.
• After that discovery: A K3.1 or K3 11111 preview appeared on the open platform, advertising one million tokens of context.

What K3.1 Would Inherit From K3

Moonshot launched K3 in its app and API in July. The July K3 announcement positioned it as the company's flagship for coding, knowledge work and complex tasks requiring extended reasoning with limited human supervision. The promised release of its weights also made access beyond Moonshot's own service a central part of its appeal.

K3 uses a mixture of experts design. In simple terms, the model contains different groups of computational components, and a routing system selects which groups help process an input. Its 2.8 trillion parameters describe the total model, while the architecture determines which parts take part in a particular calculation.

That design helps explain why the headline parameter count is only one part of the story. Users experience the combined effects of training, routing, computing resources and the software that connects the model to tools. A larger total parameter count alone does not tell a customer how quickly a coding assignment will finish or what the bill will be.

K3.1's advertised context capacity would match K3's one million tokens. The proposed upgrade therefore appears to place much of its emphasis on how the model works with information and carries out tasks, rather than on expanding the amount it can hold at once.

Reasoning Settings Could Give Users More Control

K3.1 is expected to offer three reasoning levels: Low, High and Max. These settings would give developers a way to choose how much effort the model spends working through a request. Their exact behavior and effect on billing have not been published.

A short classification task and a difficult software repair need different amounts of work. Sorting a customer message into a category may require a quick answer. Investigating a failure across several files can require repeated inspection, testing and revision. Adjustable reasoning settings could let users choose a different balance of speed and effort for each assignment.

Developers examining the platform have suggested that the levels may correspond to different allocations of computing resources and different charging arrangements. For a business processing many requests, that would make the choice an operational decision as well as a technical one.

The reported one million token window is relevant here, too. A model can receive a large collection of code or documents, but the time and computing effort needed to use that material still matter. Customers considering K3.1 will need to judge the reasoning settings alongside the size of the work they submit.

Agents and Swarm Mode Target Longer Assignments

Possible K3.1 modes include Agent, Swarm, search and batch processing. An AI agent takes a sequence of actions toward a goal, often using tools between steps. Instead of merely suggesting how to fix a program, for example, an agent might inspect files, propose a change, run a test and revise its work.

Swarm refers to several agents working together, potentially dividing a larger assignment into smaller jobs. One could search for information while another examines code and a third checks results. The proposed mode would extend that approach within Kimi, although its implementation has not been detailed.

Search and batch processing address different needs. Search helps an AI retrieve information during a task. Batch processing lets a user submit a collection of jobs, such as many documents requiring the same treatment, rather than managing every request separately.

Games, 3D coding and computer operation are also reported areas of focus. They demand more than a fluent written response: code must execute, interactive elements must behave as intended, and actions on a computer must occur in the right order. For people using AI to build or operate software, dependable execution across those steps would be a central measure of the upgrade.

Price and Speed Will Shape the Upgrade's Value

Moonshot's open platform has displayed K3.1 API pricing matching the existing K3 rates. That listing may be a placeholder, so customers still lack a settled price for the new model. The reported concerns around K3 include pricing, response speed and usage limits.

Those three issues become closely connected during autonomous work. A task that requires repeated model calls and tool use can consume more time and resources than a single answer. Usage limits can also interrupt an assignment before it is complete, leaving a person to resume or reorganize the work.

Matching K3's listed rates would therefore answer only part of the cost question. The amount of work needed to reach a useful result also affects the expense. A model that completes a task with fewer retries may be more economical for that task, even when its published rate looks similar.

Moonshot is also reportedly preparing a formal release of Kimi K2.8, which has been available as a preview for more than a month. With both K2.8 and K3.1 appearing in release discussions, developers face a product choice that will depend on the capabilities and access terms Moonshot ultimately publishes.

Competition Has Moved Since K3's July Debut

K3 entered a strong competitive position in July, debuting third on Artificial Analysis's Intelligence Index and second on the Vals AI Index. Analyst Nathan Lambert described it as the strongest open model at the time, narrowing the gap with leading American models whose internal weights remain proprietary.

Attention has since shifted following DeepSeek's release of V4.1 Flash-1017. The latest comparison described alongside the K3.1 reports places that model ahead of K3 on nearly all cited performance benchmarks. No detailed score table accompanies that claim, leaving the size of the lead unclear.

The July rankings and the newer comparison concern different moments in a fast moving market. K3 established Moonshot as a strong contender at launch; K3.1 now faces the task of showing useful progress against newer competition. For customers, the relevant comparison will include the kinds of assignments they actually run, alongside cost, latency and access limits.

The most consequential release details would connect Moonshot's proposed features to everyday results: whether an agent finishes a lengthy assignment, whether stronger reasoning justifies the wait, and whether developers can predict their spending. The registry entries and platform preview have brought K3.1 into view. A formal launch would need to turn those signals into a clearly defined product.

Key Points

• Developers found a callable "kimi-k3-1" model identifier on September 28, followed by a K3.1 or K3 11111 platform preview.
• The preview advertises one million tokens of context, matching K3's capacity.
• The 2.8 trillion parameter figure is confirmed for K3; K3.1's parameter count has not been announced.
• Expected features include Low, High and Max reasoning settings, with possible Agent, Swarm, search and batch modes.
• Displayed API pricing matches K3 but may be provisional. Speed, usage limits and task completion will also affect its value.
• Moonshot has not given a firm K3.1 release date.