Kimi K2.6 Is Out: Moonshot AI's Open-Weight Coding Model Hits 80% on SWE-Bench

Moonshot AI shipped Kimi K2.6 on April 20, 2026. The model is open-weight, meaning the weights are on Hugging Face and you can self-host. It went from invite-only preview on April 13 to generally available seven days later.

The numbers are good. On SWE-bench Verified, K2.6 scores 80.2%, up from K2.5’s 76.8%. On SWE-bench Pro, 58.6% versus K2.5’s 50.7%. Terminal-Bench 2.0, which measures long-horizon terminal task completion, went from 50.8% to 66.7%. That’s a 16-point jump on the benchmark that most directly tracks what an autonomous coding agent actually needs to do.

For context: Xiaomi’s MiMo-V2.5-Pro, which we covered last week, scores 78.9% on SWE-bench Verified and 68.4% on Terminal-Bench 2.0. Kimi K2.6 is fractionally ahead on SWE-bench and fractionally behind on Terminal-Bench. These two models are essentially trading blows at the open-weight frontier right now.

What Actually Changed from K2.5

The benchmark improvements are real, but the more interesting changes are structural.

Agent Swarm, Moonshot’s multi-agent architecture, now supports up to 300 parallel sub-agents, up from 100 in K2.5. The maximum coordinated steps grew from 1,500 to 4,000. If you’re running K2.6 in a pipeline that parallelizes coding tasks, you can throw significantly more at it before hitting limits.

Long-horizon reliability improved substantially. Moonshot ran a 13-hour continuous execution test where K2.6 independently iterated through 12 optimization strategies, made over 1,000 tool calls, and modified more than 4,000 lines of code without human intervention. That’s the kind of benchmark that doesn’t show up in SWE-bench scores but matters a lot for real autonomous coding work.

Instruction following is tighter. K2.5 had a known issue where the model would drift from architectural constraints during long sessions. K2.6 tracks them more consistently, which shows up in the OSWorld-Verified score (73.1% vs K2.5’s 63.3%).

Pricing

The official Kimi API prices K2.6 at $0.60 per million input tokens and $2.50 per million output tokens. On OpenRouter, it runs $0.74 input and $3.49 output.

For comparison: Claude Opus 4.7 is priced around $15/$75 per million tokens. Kimi K2.6 is roughly 8x cheaper on input and 10x cheaper on output for comparable coding benchmark performance. That gap matters when you’re running long agentic loops.

The context window is 262,144 tokens.

Availability

K2.6 is available through Kimi.com, the Kimi App, the official API at platform.kimi.ai, and the Kimi Code CLI. Weights are on Hugging Face under a license that allows commercial use. The Kimi Code CLI works with K2.6 out of the box.

If you were using K2.5 through the official API or Kimi Code CLI, you’re already on K2.6. Moonshot didn’t require a migration step.

Where It Fits

K2.6 is a solid option if you’re building cost-sensitive coding agents and don’t need proprietary API access. The open-weight availability means you can run it on your own infrastructure if latency or data residency matters. The 300-sub-agent swarm architecture is unusual in the open-weight space.

It’s not ahead of the frontier proprietary models on every benchmark, but it’s close enough that the price difference justifies serious evaluation. The Terminal-Bench 2.0 score is particularly notable because that benchmark is harder to game than SWE-bench and closer to what production coding agents actually need.

Sources: Kimi K2.6 Tech Blog, Hugging Face model page, OpenRouter pricing, Kilo AI K2.6 writeup