*Translation of a post from the easy-learn-ai daily digest (commit: d9b875d | 2026-04-22 AI Daily).*
On April 22, 2026, Moonshot AI did something almost unthinkable a year ago: they open-sourced the weights of Kimi K2.6 on Hugging Face, under a modified MIT license.
- 1 trillion parameters
- Mixture of Experts architecture
- Support for up to 300 parallel sub-agents
- Optimized for long-horizon coding and autonomous tasks
- Multimodal
The "Brute Force" and Restraint of a 1T MoE
K2.6 has 1 trillion total parameters, but the number of active parameters per inference is far smaller — that is the essence of MoE. Think of it as a council of 300 specialists, each with their own expertise: when a problem arises, only the most relevant few are convened, rather than dragging everyone out of bed.
These 300 "experts" can handle not just one task at a time but 300 different tasks in parallel — that's the real meaning of "300 parallel sub-agents." For complex agent workloads that require scheduling multiple tools, querying multiple data sources, and executing many independent steps simultaneously, this parallelism is a qualitative leap.
FlashKDA: Making Attention "Fly"
If K2.6 is the car body, FlashKDA is its engine.
Attention is the core computational unit of large language models — you can think of it as "the model's eyes jumping between words as it reads." Traditional autoregressive attention has a fatal weakness: to generate each new token, it must re-read all previous tokens — like re-reading your entire paper from the start after every sentence you write.
The Kimi team released FlashKDA, a CUTLASS-based implementation of Kimi Delta Attention. External tests show that on 8x MI300X, K2.6 + DFlash reaches 508 tok/s — about 5.6x faster than traditional autoregressive decoding. On H20, prefill speedups reach 1.72–2.22x.
What does that mean? Code generation that used to take half a minute to watch may now finish in seconds.
The Tipping Point from "Usable" to "Good"
A representative voice in the community: K2.6 can handle about 85% of the work of Claude Opus in real use, with browsing and vision capabilities, and is well suited to long tasks. Some users who were paying for Opus subscriptions have already switched to Kimi.
That "85%" number is interesting. It means open-source models are approaching a tipping point — for most everyday tasks, "good enough" is replacing "the best." When you can get a solution at 85% of a top closed-source model's level for zero cost (self-hosted) or very low cost (API), many people will choose it.
Nano subscriptions quickly added K2.6 and GLM 5.1. Although billed at 2x token rates, users report that "given access to the new models, these resource limits are acceptable."
The Reality of Local Deployment
Of course, a 1T model is not something everyone can run at home. K2.6 needs substantial hardware to run at full speed. But community members have already gotten it working on Macs with 256GB of RAM through various optimizations, and others run quantized versions on combinations of consumer GPUs.
This layered ecosystem — power users running it at full speed, ordinary users running trimmed versions — is precisely the greatest advantage of open-source models. Closed-source models only ever give you one "official configuration," while open-source models can be reshaped by the community into countless forms.
Why This Is a Landmark Moment for Chinese Open-Source
With DeepSeek V4 and Kimi K2.6 releasing top-tier open-source models within a month of each other, Chinese open-source AI is no longer a "follower" but a peer running in parallel.
The existence of these two models puts real pressure on closed-source pricing strategies. When users can "unsubscribe" at any time and switch to a free and strong-enough alternative, the pricing ceiling of commercial models gets redefined.
For developers, this is a golden age. You have choices, you have a fallback, and you have bargaining power.