Kimi K2: Leading Open-Source Model Now Available on Together AI

Moonshot AI's Kimi K2 is now live on Together AI, and the numbers behind it are hard to ignore. This is a trillion-parameter mixture-of-experts model — 32 billion active per token, pulling from a pool of 384 experts — trained on 15.5 trillion tokens without the instability spikes that usually plague models this size. Together credits a custom optimizer called MuonClip for that stability, which is a small technical detail with a big practical payoff: fewer failed training runs, faster iteration.

What's more interesting than the parameter count is where Kimi K2 actually wins. It tops EQ-Bench3 and creative writing leaderboards, ahead of every other model, open or closed, in those categories. On SWE-bench Verified, the standard test for whether a model can autonomously fix bugs in real codebases, it scores 65.8%, nearly double DeepSeek-V3's 38.8% and Qwen3's 34.4%. That's not a marginal gap between open-source competitors — that's a different tier.

The bigger story is how the model got good at agentic work in the first place. Instead of learning tool use passively from text scraped off the internet, Moonshot trained Kimi K2 through what it calls Large-Scale Agentic Data Synthesis — simulated interactions across hundreds of domains and thousands of tools. The result is a model that doesn't just describe what you should do with your data; it opens the file, runs the analysis, builds the chart, and writes up the findings itself. That's the difference between an assistant and an agent, and it's the gap most open models haven't closed yet.

Together AI is positioning itself as the place to actually run this thing, and the pricing makes the pitch. One dollar per million input tokens, three dollars per million output tokens — Together claims that's 60 to 70% cheaper than Anthropic's Sonnet or Opus 4, models Kimi K2 is apparently competitive with on several benchmarks. Add a 99.9% uptime SLA, multi-region hosting, and SOC 2 compliance, and the message is clear: this isn't a research toy, it's meant for production traffic.

For teams already burning budget on closed frontier APIs, an open model that scores near the top of coding and creative benchmarks at a third of the cost is the kind of thing that forces a second look at vendor contracts. Moonshot didn't just release another large model — it released one aimed squarely at the workflows enterprises actually pay for.