Recipe ·

Kimi K2.6

Moonshot AI's 1T-param MoE (32B active) built for long-horizon agentic coding (up to 13h continuous) with agent swarm scaling to 300 sub-agents. Leads SWE-Bench Pro at 58.6%, ~1/4 the cost of Claude Opus 4.6.

Techniques inside

Median research → prod

1.7y

Fastest adoption

Slowest adoption

Ingredient list

#1Test-Time Compute Scaling
Invented by Google DeepMind · 2024-08 · Velocity 1.7y
“Supports up to 13h continuous reasoning for long-horizon tasks, allocating substantial compute at inference time.”
reasoninghigh
#2ReAct (Reason + Act)
Invented by Princeton / Google · 2022-10 · Velocity 4y
“Designed for agentic coding with agent swarm scaling to 300 sub-agents, indicating ReAct-style reasoning and tool use.”
agentshigh
#3Chain-of-Thought Prompting
Invented by Google · 2022-01 · Velocity 4y
“Kimi K2.6 is built for long-horizon agentic coding with step-by-step reasoning capabilities.”
reasoninghigh
#4Mixture of Experts (Sparse MoE for LLMs)
Invented by Google · 2017-01 · Velocity 9y
“Moonshot AI's 1T-param MoE (32B active) architecture explicitly uses mixture-of-experts.”
architecturehigh

This recipe is part of the gentic.news Deployment Atlas. Every ingredient has an origin paper + evidence. Methodology is public. Dataset is CC BY 4.0.