Check Kimi K3.1 API availability, pricing and model ID status.
As of October 23, 2026, public Kimi K3.1 API access is still unconfirmed, and K3.1 is not yet available on EvoLink. Subscribe for access updates, or start with the existing Kimi K3 API.
Kimi K3.1 is the name used for a potential follow-up to Moonshot AI’s K3. It appears in public framework code, Kimi Code configuration reports and community-shared teaser material.
Status as of October 23, 2026.
| API question | Current status |
|---|---|
| Official public API | K3.1 is absent from the official model list; public access is still unconfirmed. |
| EvoLink API access | Not available yet. Register for availability updates. |
| Production model ID | Production request ID not confirmed. |
| Context window | Public API context limit not confirmed. |
| Maximum output | Unconfirmed. |
| Input and output types | Input types, output formats and file limits are unconfirmed. |
| Reasoning controls | Public API reasoning controls not confirmed. |
| Tools and structured output | Tool calling, JSON mode and schema-constrained output are unconfirmed. |
| API protocol and SDK | Protocol, base URL and SDK compatibility are unconfirmed. |
| Token and tool pricing | K3.1 rates have not been published. |
| Rate limits and streaming | Quotas, concurrency and streaming behavior are unconfirmed. |
| Account or subscription | K3.1 eligibility is unconfirmed. The developer API, membership and Kimi Code are separate services. |
Get updates on Kimi K3.1 availability and how to connect through EvoLink.
Your email is used for EvoLink model availability updates.
Discuss your integration needs: Join the EvoLink community
Follow model ID, supported parameters and pricing updates in one place. Registration is for email updates; API and beta access are handled separately.
Official documentation, public code and attributed reports answer different parts of the access question.
Official post · Community-shared reference
An October 23 community post links a teaser video from Kimi’s official X account and an associated screenshot described as showing “Model: Kimi K3.1” inside Kimi Code. The official-post link and preserved screenshot document the teaser material circulating ahead of a public API release.
Verified public code · SGLang
The SGLang commit adds a K3.1 version branch to its experimental router’s chat encoder, with regression fixtures checked against pinned reference encoders. The concrete change and its integration limits are explained below.
Official documentation
Kimi’s official model list did not include a K3.1 entry when checked on October 23.
Community discussion · Reddit
Community discussion focuses on shorter waits, fewer reasoning tokens, lower cost per completed task and coding or visual demos that hold up in daily use. Some contributors also hope for public weights. These are expectations to evaluate.
Kimi Code configuration captures offer more detail about preparation, long-context options and agent settings.
Firsthand configuration report · Lumina
Lumina published a captured Kimi Code configuration containing a model-switch rule with k3.1 as its destination. The capture also includes a warning about continuing an existing conversation after switching models. The finding suggests product preparation; a public release date and the users who will receive access remain unconfirmed.
Configuration screenshot coverage · Wccftech
September reporting described K3.1 configuration with an Extra Long context option, up to one million tokens, and Agent/Swarm flags. This concerns reported product settings; API context limits and access to product-level agent features remain separate questions.
Verified public code · SGLang
In this pinned implementation, K3.1 adds medium to the reasoning levels accepted by the K3 encoder. Prompt tags, escaping, tool rendering, segmentation and vocabulary remain shared between the two reference revisions.
KimiK3Version::K3 => &[
ThinkingEffort::Low,
ThinkingEffort::High,
ThinkingEffort::Max,
],
KimiK3Version::K31 => &[
ThinkingEffort::Low,
ThinkingEffort::Medium,
ThinkingEffort::High,
ThinkingEffort::Max,
],An eventual API exposing this setting would provide another level to evaluate for task quality, latency and token cost; the code establishes no measured gain. moonshotai/Kimi-K3.1 is a framework reference, not a confirmed request ID. This Rust excerpt describes encoder support; official API availability and accepted parameters require their own documentation.
Implementation and test fixtures · SGLang
The tests include image-format cases. The router does not fetch or decode images: it renders a text placeholder when no image prompt is supplied.
That path can match part of a prompt prefix, rather than an engine’s full image representation. It is an adaptation clue; production image inputs and file limits still need an API specification.
Turn the community’s expectations into checks on the same tasks, inputs and acceptance criteria.
Save one real repository repair task, its test suite and a clear acceptance condition. Against your current Kimi K3 baseline, record test pass rate, accepted deliverables, tool errors, retries, elapsed time, reasoning-token usage and total cost. For UI work, check appearance and interactions as well as whether the code runs.
Use documents with known answers and source locations. Score retrieval and citation accuracy, then scale the test to the context limit actually published for K3.1.
Keep the current model and fallback configuration. Before moving traffic, verify the actual route’s model ID, protocol, input types, streaming, tools, rate limits, usage and billing records.
While K3.1 access is pending, use EvoLink’s unified API gateway to establish a baseline with current models. Keep the same tasks and cost measurements for a later K3.1 comparison.
Use the existing Kimi model page to check API access, reasoning settings, tools and current pricing.
Kimi K3
Evaluate coding and agent tasks with an available model. Track completion quality, retries and total task cost.
DeepSeek V4.1 Flash 1017
Compare another current option for complex coding work. Measure quality, retries and cost per accepted result.
GPT-6.1 SolSet up authentication and requests for the existing Kimi K3 API.
Public access remains unconfirmed as of October 23, 2026. K3.1 is not yet available through EvoLink.
A production request ID has not been confirmed. Framework names and Kimi Code configuration labels are not interchangeable with an API model ID.
K3.1 rates have not been published. Input, output, caching and tools need K3.1-specific rates; K3 pricing applies to K3.
No improvement is established here. Use the same workload and the evaluation checks above to compare quality, time and total task cost.
The public discussion includes code that renders interactive scenes. That differs from native video generation or production-ready 3D assets, neither of which is confirmed for K3.1.
No public weight release or license is confirmed. A framework repository reference does not establish downloadable weights or usage permission.
Moonshot controls its own products and API. EvoLink’s gateway integration is a separate availability decision.
Both API limits remain unconfirmed. Context capacity and maximum output are separate limits; the configuration report above does not establish an API contract.
The public API’s input types, image/video support, output formats and file limits are unconfirmed. The image-test section explains the framework evidence.
Public API tool calling, JSON mode and schema constraints are unconfirmed. Agent/Swarm product settings do not specify API fields or limits.
K3.1 SDK compatibility, base URL and supported fields are unconfirmed. The linked K3 guide covers the current K3 setup.
The developer API, membership and Kimi Code are separate services. K3.1 eligibility is unconfirmed; EvoLink’s form registers email updates.
Kimi K3 is available now. Its linked model page provides current access, input, tool and pricing details for a baseline and fallback.
Public API and EvoLink access dates are unconfirmed. Register for updates when access details become available.