Kimi K2 Guide: The 0711 and 0905 Snapshots in 2026

Kimi K2 is the model the rest of Moonshot's current line is built on. K2 Thinking, K2.5, K2.6 and K2.7 Code all share its basic shape: a trillion parameters in total, 32B active per token. Understanding K2 is a shortcut to understanding all of them.

It is also fifteen months old, which in this field is a long time. This guide covers what K2 is, the two snapshots that share its name, and the narrow set of reasons to still call it directly.

What Moonshot shipped in July 2025

Moonshot released kimi-k2-0711-preview on 11 July 2025. The model card describes a mixture-of-experts model with roughly 1T total parameters and 32B active, 384 experts with eight selected per token plus one shared, 61 layers and MLA attention. It was pretrained on 15.5T tokens using an optimiser Moonshot calls MuonClip, its adaptation of Muon with additions to keep training stable at that scale.

Two variants were published: Kimi-K2-Base, the foundation model for teams who want to fine-tune, and Kimi-K2-Instruct, the post-trained model for chat and agent use. Moonshot described the Instruct model as reflex-grade, meaning it answers without a long hidden reasoning phase. That is the defining trait of K2 compared with almost everything that followed it.

If the total-versus-active split is new, mixture of experts explained covers why 32B active is the number that drives speed and cost, not the trillion.

0711 and 0905 are different models under one name

Moonshot released kimi-k2-0905-preview on 5 September 2025. It is easy to treat the date suffix as a minor patch, and that would be a mistake. The 0905 model card lists real changes:

• Context doubled from 128K to 256K tokens.
• SWE-bench Verified rose from 65.8 to 69.2 on Moonshot's harness.
• SWE-bench Multilingual rose from 47.3 to 55.9.
• Terminal-Bench rose from 37.5 to 44.5.

Moonshot also reported better frontend output. All of these are vendor figures: the 0905 numbers are means over five runs, mostly on Moonshot's in-house harness (Terminal-Bench used Terminus-2), so treat them as direction rather than precise rank.

The practical consequence is about which identifier you call. On OpenRouter, moonshotai/kimi-k2 maps to the 0711 snapshot and is listed at $0.57 input and $2.30 output per million tokens, while moonshotai/kimi-k2-0905 is listed at $0.60 and $2.50, at the time of writing. For a few cents more you get twice the context and the stronger coding results. Unless you have a regression test proving 0711 behaves better on your task, 0905 is the one to use. Pinning model versions covers why you should name the snapshot explicitly either way.

What the modified MIT licence asks

Both the code and the weights are released under what Moonshot calls a Modified MIT licence. It is MIT with one addition: if your product passes 100 million monthly active users or $20 million in monthly revenue, you must display the Kimi K2 name prominently in its interface.

For most teams that threshold is irrelevant, and the licence behaves like MIT. It is still not MIT: the display requirement is an extra condition that plain MIT does not have, and it applies to derivative works too. Legal review should read it as written. Open-weight model licences compared puts it next to the alternatives.

A non-thinking model in a reasoning-model world

K2 does not emit a long reasoning trace before it answers. Its successors mostly do. That is a disadvantage on hard multi-step problems and an advantage everywhere else.

A reasoning model bills you for its thinking tokens, and on simple work that thinking is overhead: classification, extraction, rewriting, short tool calls where the right action is obvious. On that kind of traffic a reflex model returns sooner and spends less. The token cost of reasoning models covers how large that overhead gets.

Moonshot recommends temperature 0.6 for K2 Instruct. If you are moving prompts over from a reasoning model that wanted temperature 1.0, that is one of the settings to change.

Should you run K2 in 2026?

Mostly, no, and the reason is the rest of the family. K2.6 keeps the same 1T/32B shape with a 256K window and adds native vision. K2 Thinking added long-horizon reasoning on the same base, and K2.7 Code specialises it for programming. The Kimi K2.6 guide is the place to start for a current general model, and the Moonshot family overview maps the whole line.

There are three reasons to keep K2 in play:

1. An existing pipeline that works. If K2 0905 passes your evaluation and costs less, switching for the sake of novelty is risk with no payoff.
2. Fine-tuning from a base model. Kimi-K2-Base gives you a non-instruction-tuned starting point at this scale.
3. Reflex workloads. High-volume, simple requests where reasoning tokens are pure cost.

If any of those apply, test K2 0905 against K2 Thinking and K2.6 on the same twenty tasks, and compare cost per completed task. Otherwise, start your evaluation one or two generations later.