Kimi Rolls Out K3.1 Pro in Kimi Code, With a Faster Variant That Costs Twice the Quota
Kimi has rolled out K3.1 Pro across Kimi Code, according to the product's model documentation, published on 23 October 2026. It is the first model in the new K3.1 series, which Kimi describes as its most capable 3T-class model to date, with 3T parameters. Parameters are the learned weights that determine how a model responds, and 3T means roughly three trillion of them.
The company is also releasing Kimi K3.1 Pro Highspeed, which Kimi says has the same model capabilities as K3.1 Pro. Kimi puts its speed at roughly two to three times that of the standard model. The trade-off is quota: according to the same documentation, Highspeed consumes about twice the quota of the regular K3.1 Pro. Access to Highspeed requires an Allegretto or Pro plan or above.
The two models are addressed by the identifiers k3.1-pro and k3.1-pro-highspeed. Both accept image and video input for multimodal requests, meaning requests that combine text with other media. The Kimi Code model documentation lists K3.1 Pro with a 1M-token context window, given as up to 1,048,576 tokens for higher-tier members. A token is a fragment of text or code, roughly three-quarters of an English word, and the context window is the total the model can consider in one session.
Thinking mode is mandatory on both models. K3.1 Pro models support only thinking mode, and any request with thinking disabled is rejected. Integrators who previously switched reasoning off for individual requests will need to remove that parameter from their clients, or expect errors.
The rollout sits on top of the K3 generation. Kimi describes K3 as the world's first open 3T-class model, and lists its flagship K3 at 2.8 trillion parameters in the Kimi membership guide. Moonshot AI describes K3 as natively multimodal, with a 1M-token context, and built for long-horizon coding, knowledge work, and deep reasoning. A fixed-context variant, k3-256k, carries a 256K limit. K3 accepts three effort levels, low, high, and max, as shown in Kimi's third-party tools documentation for Codex. Using k3 with up to 1M context requires an Allegretto or Pro plan or above, according to Kimi's error reference.
The figures that matter most to a platform team are mostly the ones that trade against each other. A 1M-token window lets an agent hold a large repository, long build logs, or a lengthy design thread in one session, but Kimi ties the full window to its higher plan tiers. Highspeed cuts latency for interactive work, yet at twice the quota cost, so the effective price of a fast session depends on how much of a plan's allowance a team's workload consumes. Mandatory thinking means every request carries reasoning tokens, which affect both latency and billing even when the visible output is short.
In my view, the more useful question for engineering leads is not the parameter count but the cost per completed task under their own quota and latency constraints. Vendor speed multipliers and context ceilings are measured under conditions that rarely match a team's codebase. Trials on representative tasks will answer that more reliably than the headline figures.
The broader context here is a large open-weight-lineage model with multimodal input and a very long context, now callable from a tool developers already use. Kimi's framing, built for long-horizon coding and deep reasoning, points at multi-hour agentic sessions as the target. Whether those sessions become routine will depend on how the tiered quotas and mandatory thinking settle in practice, and on how many teams judge that a 1M-token window and a faster variant justify the plan needed to reach them.
