@KKaWSB: Moonshot just open-sourced Kimi K2.6—4,000 tool calls in one 12-hour session, 300 sub-agents in parallel building a full codebase. SOTA on SWE-Bench Pro, BrowseComp, HLE and more, ties Claude Opus 4.6 and G…
Summary
Moonshot has open-sourced the Kimi K2.6 model, supporting 4,000 tool calls in a single session and 300 parallel sub-agents, achieving SOTA on benchmarks like SWE-Bench Pro and claiming performance on par with Claude Opus 4.6 and GPT-5.4.
View Cached Full Text
Cached at: 04/21/26, 08:27 AM
Moonshot just open-sourced Kimi K2.6 — 4,000 tool calls in a single session for 12 straight hours, 300 sub-agents in parallel building an entire codebase.
SOTA on SWE-Bench Pro, BrowseComp, HLE and more benchmarks, tying Claude Opus 4.6 and GPT-5.4.
Open-source models have caught up to top-tier closed-source — the only reason you thought they weren’t good enough is that you hadn’t tried the latest.
Similar Articles
Kimi K2.7 Code: 1T MoE, $0.95/M tokens, MIT license, beats Opus 4.8 on MCP tool-calling
Moonshot AI 发布了专注于编程的开放式权重模型 Kimi K2.7 Code，拥有1万亿参数和384个专家，性能在MCP工具调用上超越Opus 4.8，成本仅为十分之一。
@akshay_pachaar: Kimi K2.6 raises the bar for open-source models. Moonshot released it yesterday, and for the first time, an open-weight…
Moonshot's open-weight Kimi K2.6 matches Claude Opus 4.6 on key agentic benchmarks while costing significantly less.
@heyshrutimishra: OpenClaw users are gonna love it Finally an open source model that beats Opus 4.6 on SWE-Bench It's Kimi K2.6, it runs …
Kimi K2.6 open-source model surpasses Opus 4.6 on SWE-Bench, supporting 12+ hour autonomous coding sessions with 4,000+ tool calls.
@QuixiAI: @Kimi_Moonshot K2.6 running on my mi300x, 56 tps (single request). I will run a throughput test
Kimi K2.6 achieves 56 tokens per second on a single MI300X GPU; user plans further throughput benchmarking.
@sanbuphy: K2.6 successfully downloaded and deployed the Qwen3.5-0.8B model locally on a Mac, using the niche Zig language to implement and optimize inference, demonstrating the new model’s generalization ability. After 4,000+ tool calls and 12+ hours of continuous operation, K2.6 iterated 14 times…
K2.6 successfully downloaded and deployed the Qwen3.5-0.8B model locally on a Mac, using the niche Zig language to implement and optimize inference, demonstrating the new model’s generalization ability. After 4,000+ tool calls and 12+ hours of continuous operation, K2.6 iterated 14 times, boosting throughput from ~15 tokens/s to ~193 tokens/s, ultimately achieving 20% faster inference than LM Studio.