- Moonshot AI open-sourced Kimi K2.6 on April 20, 2026, according to its own help center. It is a 1T-parameter mixture-of-experts model (32B active) with a 256K context, released under a Modified MIT License.
- Moonshot's K2.6 model card and launch blog say Agent Swarm scales to 300 sub-agents and 4,000 coordinated steps. The "1,000 agents" figure in early-October news items doesn't appear in those materials. The 10-hour SysY compiler build is a Moonshot claim reported in April.
- Kimi K2.6 costs $0.95 per million input tokens and $4.00 per million output tokens on Moonshot's API. Moonshot's newer flagship, Kimi K3 (released July 2026), now powers Kimi's Agent Swarm.
Contents
In early October 2026, a series of short news items described Kimi K2.6 from Moonshot AI as a new open-source model that can coordinate "up to 1,000" agents and build a compiler in 10 hours. Moonshot's own records don't support the timing. Moonshot open-sourced Kimi K2.6 on April 20, 2026, and three months later replaced it as its flagship with Kimi K3. This piece covers what K2.6 is, what Moonshot actually claims for it, what it costs, and where it fits among today's open-weight models.
This story is based on Moonshot's Hugging Face model card, launch blog, help center and API pricing page, and on April 2026 coverage of the launch, checked October 12, 2026.
The short version
- Released: April 20, 2026. Moonshot's help center says K2.6 "launched as open source" on that date.
- Size: 1 trillion total parameters, 32 billion active per token (mixture-of-experts), with a 256K-token context and a MoonViT vision encoder.
- License: Modified MIT. Very large products must show "Kimi K2.6" in their interface.
- Agent Swarm: up to 300 sub-agents and 4,000 coordinated steps, per Moonshot. Not 1,000.
- API price: $0.95 per million input tokens, $0.16 cached and $4.00 output.
- Newer model: Kimi K3, released in July 2026, now powers Kimi's Agent Swarm.
When was Kimi K2.6 released?
Moonshot's Agent Swarm help page gives the dates. Kimi K2.5 introduced Agent Swarm on January 27, 2026, and "Kimi K2.6 launched as open source" on April 20, 2026. Coverage from the time agrees: MarkTechPost reported the open-sourcing on April 20, and VentureBeat wrote about it on April 21.
The October items, several from NewsBytes between October 2 and October 4, say Moonshot "just announced" the model. They link to no Moonshot announcement. We found no new K2.6 release, version or weights update from Moonshot in October. The repackaged claims (SysY compiler, five-day agent run, "Claw Groups") all come from April launch materials.
What is Kimi K2.6?
According to the model card, K2.6 is an "open-source, native multimodal agentic model":
- 1T total and 32B activated parameters, with 384 experts (8 selected per token plus 1 shared) across 61 layers
- MLA attention and a 160K vocabulary
- a 256K context window (262,144 tokens in evaluations)
- the same architecture as K2.5, so existing vLLM, SGLang and KTransformers deployments carry over
- the "same native int4 quantization method" as Kimi-K2-Thinking
That's a multi-GPU server model. Moonshot recommends inference engines but doesn't give a minimum hardware spec on the card.
Does Kimi K2.6 really run 1,000 agents?
Not according to Moonshot's published materials. Both the model card and Moonshot's K2.6 launch blog say Agent Swarm scales "to 300 sub-agents executing 4,000 coordinated steps", up from 100 sub-agents and 1,500 steps in K2.5. The help center also gives 300 as the current limit. The only "1,000" in the K2.6 blog is a run with "over 1,000 tool calls."
The October items attribute a line to founder Zhilin Yang about "orchestrating 100 or even 1,000 sub-agents in parallel," without a link or date. Even if he said it, that describes an ambition, not a shipped limit.
The other headline claims are Moonshot's own and were reported in April:
- SysY compiler: VentureBeat reported on April 21 that, according to Moonshot, K2.6 "built a full SysY compiler from scratch in 10 hours" and passed all 140 functional tests. Moonshot compared that to a four-engineer team working for two months. SysY is a small subset of C used in Chinese university compiler courses. The claim doesn't appear in the K2.6 blog we checked, and no one has reproduced it independently.
- Five-day run: the blog says an agent backed by K2.6 "operated autonomously for 5 days," handling monitoring and incident response for Moonshot's RL infrastructure team.
Our explainer on what "agentic" really means covers why long unattended runs say more about the harness and task than about the model alone.
How good is it? (vendor-reported)
The model card lists Moonshot's own scores against models from spring 2026. Among them: 80.2 on SWE-Bench Verified, 58.6 on SWE-Bench Pro and 66.7 on Terminal-Bench 2.0, each shown next to spring-2026 competitors. Some competitor scores were re-run by Moonshot, and settings vary by benchmark. Every comparator on the card has since been superseded, so these numbers are useful for history and little else.
How much does Kimi K2.6 cost?
| Model (Moonshot API) | Input (cache miss) | Input (cache hit) | Output | Context |
|---|---|---|---|---|
| kimi-k2.6 | $0.95 | $0.16 | $4.00 | 262,144 |
| kimi-k3 | $3.00 | $0.30 | $15.00 | 1,048,576 |
Prices are per million tokens. Last verified: October 12, 2026 (Moonshot platform pricing page)
The weights are free to download. The license is MIT with one condition: a commercial product or service with more than 100 million monthly active users, or more than $20 million in monthly revenue, must "prominently display 'Kimi K2.6'" in its interface. The Kimi K3 license goes further: it requires model-as-a-service providers above $20 million in 12-month revenue to sign a separate agreement. Our open-weight licenses comparison covers how these custom terms compare.
Where does K2.6 fit now?
K2.6 is no longer Moonshot's flagship. Kimi K3, released in July 2026, is a 2.8T-parameter model (104B active) with a 1M context, according to its model card, and the help center says Agent Swarm is "currently powered by Kimi K3." K2.6 is the cheaper option: about a third of K3's list price on Moonshot's API. Our frontier API vs open model cost comparison prices K3 on third-party hosts. For a recent open model with a permissive license, see DeepSeek-V4.1-Flash.
What's confirmed, reported and uncertain
- Confirmed (Moonshot): the April 20, 2026 open-source release, the 1T/32B architecture, the 256K context, the Modified MIT license, the 300-sub-agent and 4,000-step Agent Swarm limits, API prices, and Agent Swarm now running on K3.
- Reported (VentureBeat, April 2026): the 10-hour SysY compiler build and its 140 tests, as Moonshot described them.
- Unverified: the "1,000 agents" figure and the Zhilin Yang quote in October news items, and any October 2026 K2.6 update.
About this storyBased on the sources linked below. Editorial standards
