Moonshot AI released Kimi K3.1 Pro on October 23, 2026, alongside a Highspeed variant that it says delivers two to three times the speed. In Kimi Code, Highspeed uses about twice the quota of regular K3.1 Pro. The dated release notes confirm the rollout, and the model configuration guide explains the quota comparison and limits of the speed increase. No benchmark measurements are supplied for this report.
Moonshot's announcement lists availability across Kimi's apps, Kimi Work, Kimi Code and the API. Full model weights are promised by November 4, 2026.
Faster output leaves tool time unchanged
Highspeed only speeds up model output. Tool calls and script execution are unaffected, according to the configuration guide. A coding task that spends most of its time running commands therefore has less to gain than one that spends most of its time waiting for generated text. The advertised speed cannot be read as a two-to-threefold reduction in total task time.
That distinction gives developers a practical way to judge the quota premium: check where the agent spends its time. For a workflow dominated by builds or tests, using more quota for faster generation may save little elapsed time. This follows from the documented limit; it is not a measured result from this report. Moonshot describes the two variants as having the same model capabilities.
Check access and thinking settings before switching
The launch notes put regular K3.1 Pro access on Plus or Moderato plans and above. Highspeed requires Pro or Allegretto and above. Those higher tiers also unlock the regular model's context window of up to one million tokens in Kimi Code. Model capacity and subscription access are separate limits.
Both K3.1 Pro variants require thinking mode. Disabling thinking returns a 400 error. The supported effort levels are low, high and max, with max as the default. An integration configured to turn thinking off needs that setting changed before it can use these models.
Kimi also recommends starting a fresh session when moving to K3.1 Pro. Its guide warns that switching models or thinking effort invalidates the context cache, requiring another prefill and increasing token use. Keep the model and effort fixed within each session to avoid that additional consumption when evaluating the tiers.
The next dated milestone is the promised weights release by November 4. For hosted Code users, the immediate question is how much total task time Highspeed saves for its additional quota. The launch's speed claim alone cannot answer it.