Kimi K3.1 Pro launches with 1M context; open weights are still due
Kimi K3.1 Pro is live via apps and API with up to 1M tokens and a Highspeed tier. Full downloadable weights are only promised by November 4.
Kimi K3.1 Pro became available through Kimi products and the API on October 23, 2026. Moonshot advertises up to one million context tokens and a Highspeed option that it says is two to three times faster. Full model weights are promised by November 4, so the hosted launch is not evidence that self-hostable weights are already released.
CLAIM-LEVEL EVIDENCE
What is verified and what is not
Vendor reporting is not independent verification.
- Primary source
Hosted launch and plans
The official Kimi Code record lists the October 23 launch and tier-specific access.
- Vendor claim
Parameters, context and speed
Moonshot reports a 2.8T-total/104B-active MoE, up to 1M context and approximately 2–3× Highspeed.
- Open question
Downloadable weights
Full weights are expected by November 4, 2026; no completed release is established in this report.
- Editorial assessment
Security and benchmark claims
Cyber-filtering and model-comparison scores require independent operational assessment.
Primary-source and scope-limited editorial review
Open-weight models & agents: technical details and analysis
The most consequential distinction in Moonshot AI's Kimi K3.1 Pro announcement is between a model that customers can use and one they can download. The company made the new model available across its apps and API while placing the full-weight release in the future, by November 4, 2026. K3.1 Pro is described as a heavily post-trained successor to Kimi K3 in the roughly three-trillion-parameter class. Buyers should treat hosted access, open-weight availability and eventual licensing as three separate claims.
Moonshot specifies a mixture-of-experts design with approximately 2.8 trillion total parameters and 104 billion active parameters. It advertises a context window of up to one million tokens, potentially useful for lengthy codebases, multi-document analysis and tool-using agents with substantial histories. A maximum window is not an assurance that every subscription unlocks it. Kimi Code's release documentation explicitly distinguishes membership tiers, with higher plans required for the full context configuration and the separate Highspeed model.
The Highspeed variant is advertised as delivering about two to three times the response speed at equivalent model quality. Coding, computer-assisted 3D production, finance research and long-horizon agent tasks form the manufacturer's primary use cases. For a production agent, headline benchmark scores are not enough: success also depends on permission boundaries, tool reliability, recovery from errors and whether a reviewer can understand each consequential action. Any evaluation should measure representative end-to-end tasks, not just one curated benchmark.
Moonshot also says it removed cyber-related material from training stages to limit offensive capability in its planned open-weight release. That is a supplier-described risk-reduction decision, not independent proof that cyber abuse is impossible. The same evidence standard applies to comparative benchmark charts: figures produced or selected by the model maker require scrutiny of tool harnesses, model versions, fallback behavior and scoring rules before they can guide procurement.
Availability differs across channels. The company points to Kimi, Kimi Work, Kimi Code and its API, while the Code release notes specify separate plan entitlements. Enterprise organizations may need administrator enablement. Data handling, region commitments, logging, API retention settings and contractual protections remain distinct matters to verify with the provider. Once the weights actually appear, their license, exact checkpoint and practical hardware requirements will have to be checked before anyone can responsibly claim independent deployment.
Our editorial assessment is that K3.1 Pro matters because it may widen the competitive field for organizations that want more infrastructure choice. Yet an announced open-weight roadmap is not an artifact, and an API is not a private instance. Testing the hosted edition now makes sense for teams with suitable data controls; a self-hosting decision should wait for the actual weights and terms to be published and independently inspected.
Sources and further reading
- Moonshot AI – Kimi K3.1 Pro technical announcement ↗
- Kimi Code – Official release and tier requirements ↗
No independent product, tenant or laboratory testing was conducted.
Frequently asked questions
Can I download the full Kimi K3.1 Pro weights now?
No. The manufacturer promises the complete weights by November 4, 2026; the current hosted and API access is different.
Does every Kimi subscription include the 1M-token window?
No. Official Kimi Code notes distinguish plan tiers, with higher memberships needed for the full window and Highspeed.
Do vendor benchmark scores establish deployment readiness?
No. Controlled application trials, vendor contracts and reproducible benchmarks are needed to assess a particular use case.