s/lilianwengOPEN MODEL•Apr 21
152
votes
5.1k
seen
Kimi K2.6 runs 300-agent swarms through 4,000-step workflows
An open-weight coding model just landed everywhere at once and is claiming top-tier benchmark wins against closed systems. Kimi K2.6 showed up on Hugging Face, Arena, and now HuggingChat within a day, pushing a long-horizon coding setup with a 256k context and agent swarms that run up to 300 parallel subagents through 4,000-step workflows.
• 58.6 on SWE-Bench Pro, 54.0 on HLE with tools, 76.7 on SWE-Bench Multilingual
• 32B/1T MoE architecture with 256k context and built-in agent workflows
Benchmarks put it at or near open-source SOTA, but early hands-on use says the gap to top closed models still shows up in rough edges during real tasks despite strong coding and reasoning.
• 58.6 on SWE-Bench Pro, 54.0 on HLE with tools, 76.7 on SWE-Bench Multilingual
• 32B/1T MoE architecture with 256k context and built-in agent workflows
Benchmarks put it at or near open-source SOTA, but early hands-on use says the gap to top closed models still shows up in rough edges during real tasks despite strong coding and reasoning.
1 comment
Apr 21