Moonshot introduces Kimi K2.6, an open-weight model that it says shows strong improvements in long-horizon coding tasks, available under a modified MIT License
We are open sourcing our latest model, Kimi K2.6, featuring state-of-the-art coding, long-horizon execution, and agent swarm capabilities.
Kimi AI
Context & Ripple Effects
K2.6 follows Moonshot’s K2 and K2.5 releases, which emphasized a large MoE design and the ability to coordinate large agent swarms. The release therefore extends an established effort to make Kimi a platform for coding and multi-step task execution rather than a single-turn model.
Later coverage of K2.7-Code focuses on reducing reasoning-token use versus K2.6, while Moonshot’s planned K3 weight release suggests the company is pairing rapid model iteration with continued weight distribution.
First-order effects
Developers can obtain and adapt K2.6 weights under Moonshot’s modified MIT License, rather than relying solely on a hosted Kimi offering.
Moonshot puts a new claimed capability benchmark—long-horizon coding and agent-swarm execution—into its open-weight product line, subject to users validating those claims in their own workloads.
Second-order effects
Open-weight access gives coding-tool builders and enterprise AI teams another candidate for self-hosted or customized agentic workflows, increasing pressure on model vendors to differentiate on performance, operating cost, support, or license terms.
The subsequent emphasis on lower reasoning-token usage in K2.7-Code indicates that capability alone will not settle adoption: inference efficiency becomes a key comparison point for long-running coding agents.
Third-order effects
If Moonshot continues releasing increasingly capable weights, competition in agentic coding may shift from exclusive model access toward implementation advantages: evaluation, orchestration, deployment, and the economics of running agents at scale.
Modified open licenses make access governance part of model procurement: organizations will need to assess permitted use and distribution terms alongside coding performance and token efficiency.
The trend: This is part of the shift toward open-weight models competing for agentic coding workloads through a combination of long-horizon capability, deployment flexibility, and inference efficiency.
dear god lol - new kimi model is a fucking beast. GPT 5.4 level coding, 76% cheaper than opus 4.7 and 100% open source / free to use, i mean look at this: > kimi k2.6 can code continuously for 12 hours straight, run 300+ agents in parallel from a SINGLE prompt. how the fuck
Kimi K2.6 wrote an inference engine for Qwen3.5 0.5B in Zig and managed to beat LM Studio's token per second by 20%, running for 12 hours and with 4000+ tool calls [image]
@johnbuilds This doesn't make any sense? Ant has never pushed “European style” regulation. We've called for safety testing and transparency for frontier labs, plus strong export and distillation controls to help maintain the US advantage. Re: Kimi in particular: https://www.anthr…
BTW, I vibe coded this LLM inference engine example in the official blog using Kimi K2.6 on my laptop😘. I choose to use zig, not because it is easy, but because it is hard. I've never written any zig and metal code in my entire life, and I can just build whatever I imagine [image…
I knew K2.6 will be this strong, but I had doubts they'll open source it. Because this *is* it, the Big League, the frontier. This is our standard test of building a basic HTML ASCII roguelike, then evolving it in three steps (step 2: topdown/isometric, step 3: full voxel 3d [vid…
I'm not usually a fan of OSS models, but this new Kimi release looks pretty great (on the surface) and is priced extremely well. I plan to try it later today... how has it been so far for you? [image]
Kimi K2.6 is open-source! Open-source SOTA: - SWE-Bench Pro: 58.6 - beats GPT-5.4 (xhigh) and Claude Opus 4.6 (max effort) I realize Kimi is shipping faster and faster. An S-tier open-source model team. Keep the open-source vs. closed-source AI gap at 6 months! 💪 [image]
Holy smokes, “state-of-the-art results on several coding and tool-use benchmarks” Open Source! Kimi is cooking! -54% HLE w tools -58.6% SWE Bench Pro -66.7% Terminal Bench tl;dr Moonshot AI released Kimi K2.6, an open-source model claiming state-of-the-art results on several [ima…
this part of the KIMI K2.6 launch blog is insane: > it deployed Qwen3.5-0.8B model locally on a Mac. > coded and optimized its inference in Zig > (never knew you could do that) > improved throughput from ~15 to ~193 tokens/sec > made it 20% faster than LM Studio > did 4,000+ [ima…
the anthropic lobbyist argument should be discarded and laughed at. china open source models are within spitting distance of anthropic (according to evals). we need to accelerate not apply europe style regulation to our innovators
kimi k2.6 matches opus 4.6 in benchmarks and licensing didn't change, same modified-mit open-source license! this looks like a big upgrade for long-horizon agents!
Not comparing this against Opus 4.7.. is kinda chicken of them Nonetheless an Open Source model scoring 58.6 on SWE PRO is the highest we've ever seen [image]
kimi K2.6 vs K2.5, mythos, opus 4.7, and cursor composer 2 (based on K2.5) on every benchmark i could find tl;dr: it's a really really good model [image]
life when you discover an open-source model that runs 300 parallel agents, executes for 12+ hours straight, beats GPT-5.4 and opus 4.6 on multiple benchmarks... and the weights are on huggingface [video]