It is the short form of Kimi K2.6. Kimi's help center dates the K2.6 open-source release to 20 April 2026 and ties it to a larger Agent Swarm: up to 300 subagents and more than 4,000 tool calls in one task.

What is Kimi K2 0905?

The 0905 instruct update of Kimi K2: 256K context and a coding-focused weight, documented on the K2 announcement and on the Hugging Face card Kimi-K2-Instruct-0905. It is an earlier checkpoint than K2.6.

Is K2.7 Code included in the $19 plan?

The $19 Moderato plan is a membership for Kimi's products, including Kimi Code under that plan's limits. Calling kimi-k2.7-code on the API is a token bill. The K2.7 Code page describes those as alternatives.

Can I download the weights?

The 0905 instruct card is published on Hugging Face under a Modified MIT license, in block-fp8. Running a 1T-parameter MoE is a cluster job, not a laptop chat install. The app and the API are the practical ways to use it.

Kimi K2, K2 0905, K2.6, and K2.7 Code

Which K2 name is a checkpoint, which is a code model, and which is a membership.

K2 is a model family, not a plan name

Kimi K2 is Moonshot AI's mixture-of-experts model line. The launch post describes K2 as 1 trillion total parameters with 32 billion active, aimed at knowledge, math, coding, and agent-style tool use. Moonshot open-sourced a base checkpoint and an instruct checkpoint, and pointed API users at platform.moonshot.ai. A search for k2 6, Kimi K2 0905, or Kimi K2.7 Code is someone trying to tell those checkpoints apart.

The membership names (Moderato, Allegretto, Allegro, Vivace) are not model names. You can be on a plan and still choose which K2-generation the product runs. You can also skip the app and call a model id from the API.

What K2 0905 refers to

Kimi's K2 post has an update marked 0905: a newer weight with stronger agentic coding and a 256K context, up from the earlier 128K window. The Hugging Face model card for moonshotai/Kimi-K2-Instruct-0905 repeats the architecture: mixture-of-experts, 1T total, 32B activated, 384 experts, 8 experts selected per token, 61 layers, vocabulary 160K, context length 256K, attention MLA, activation SwiGLU. Weights and code on that card are under a Modified MIT license.

So Kimi K2 0905 is the September 2025 instruct update of K2, not a separate chatbot and not the 2026 swarm release. If a host still offers the id kimi-k2-0905, you are calling that checkpoint. Later products in the Kimi app may default to a newer generation even when people still type 0905.

K2.6 and K2.7 Code

Kimi's Agent Swarm help page says that on 20 April 2026 Moonshot released and open-sourced Kimi K2.6, and that the swarm on K2.6 can deploy up to 300 subagents and run more than 4,000 tool calls in one task. The page later says Agent Swarm in the product is powered by Kimi K3 (K3 Swarm). Treat K2.6 as a model generation and K3 Swarm as the product label Kimi uses for the parallel mode. The swarm page stays on the product behavior.

Kimi K2.7 Code is the coding model on Kimi's K2.7 Code pricing page. That page lists kimi-k2.7-code at $0.19 per million input tokens on a cache hit, $0.95 on a cache miss, and $4.00 per million output tokens, with a 262,144 token context. The highspeed id doubles those three rates. A search for Kimi K2.7 Code is this model and its API card, which is also what you would point a Python workflow at if you are calling the API rather than chatting.

Kimi K2

K2 is a model family, not a plan name

What K2 0905 refers to

K2.6 and K2.7 Code

Questions

Keep reading

Sources