Moonshot AI announced Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed on October 23, 2026, introducing a 1-million-token context window and a claimed 2–3× speed advantage for Highspeed.
October 23 brought services, not a verified weight download
Moonshot’s launch post lists availability through Kimi.ai, Kimi Work, Kimi Code and the Kimi API. It separately promises full model weights by November 4, 2026. Vercel’s independently dated October 23 changelog confirms both variants were available through AI Gateway that day and repeats the future weights plan. These are hosted releases, not merely a teaser or a preview announcement; neither announcement establishes that checkpoint files were downloadable.
As of our November 1, 2026 check, Moonshot’s official Hugging Face catalog listed 19 models, including Kimi K3, but no Kimi K3.1 Pro checkpoint. We therefore could not verify downloadable K3.1 weights. November 4 remains a promised deadline, not an accomplished release, and the earlier K3 repository is not a substitute.
The operational implication is straightforward: evaluate hosted access now if appropriate, but do not approve a self-hosted K3.1 deployment on the assumption that the required artifacts already exist. A future checkpoint release needs its own review of license terms, files and serving support.
What Pro and Highspeed change for agent builders
The direct API identifiers are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot’s quickstart documents always-enabled thinking, reasoning-effort settings of low, high and max, with max the default, plus streaming and tool calling. For multi-turn tool workflows, it instructs developers to preserve the complete assistant message rather than only its final content.
That makes a migration more than a model-name swap. Our recommendation is to replay a representative agent trace, verify that tool results and assistant state survive each turn, and explicitly record reasoning effort. Otherwise, a speed comparison can accidentally compare different amounts of reasoning rather than the two serving options.
Moonshot’s 2–3× Highspeed figure is a vendor claim, not a measured result from Nerova. Vercel positions the faster variant for interactive use. For a customer-facing assistant, test time to the first useful answer and total task completion, including external tools; for background work, prioritize successful completion within the required deadline. Faster generation alone does not establish a proportionately faster business workflow.
Capability claims need their operating conditions
Moonshot describes K3.1 Pro as a post-trained K3 with substantially more post-training compute. It reports 65.7% on Terminal-Bench 4.0 and 61.2% on its internal Kimi Code Bench 3.0 Long-horizon. Its footnotes say results generally use maximum reasoning effort and identify the Claude-Code harness for those coding evaluations. These are attributed launch results, not our reproduction or a guarantee for a company’s repository.
For a practical pilot, pair a code-change task with executable acceptance tests and a document-analysis task with source checks. Keep the same prompts, tools and review criteria across variants. Treat the large context window as input capacity to evaluate, not proof that every relevant fact will be retrieved correctly.
The next milestone is evidence of an actual release
Before changing a deployment plan around November 4, look for an official checkpoint announcement, inspect the downloadable files and license, and confirm which model and artifacts were released. Do not assume that a hosted Highspeed identifier implies a separate downloadable Highspeed checkpoint.
Our assessment on November 1: the October 23 event justifies a controlled hosted pilot, while self-hosting remains a separate, unverified release milestone. Keep those decisions separate so a service launch does not silently become a promise of local deployment.