Kimi K3 on OutcomeDev, Day One. Frontier Agentic Coding with Open Weights
Moonshot AI just shipped the largest open-weight-bound model ever. Here is what the benchmarks actually say, where it wins, where it doesn't, and how to run it on OutcomeDev today.
Moonshot AI released Kimi K3 today, and it is the most serious open-weight challenge to the frontier labs yet: a 2.8 trillion parameter Mixture-of-Experts model with a new attention architecture, a 1 million token context window, and benchmark numbers that put it inside the top three model families in the world. It is available on OutcomeDev now. Pick it in the model selector, bring your Moonshot key, and it runs.
What K3 actually is
K3 is a generational jump, not a fine-tune. The headline facts:
- 2.8 trillion total parameters in a Mixture-of-Experts design, the largest open-weight-bound model ever announced. Moonshot has promised the full weights by July 27, continuing the Modified MIT pattern it set with the K2 family.
- Kimi Delta Attention, a new attention architecture aimed at long-horizon agent workloads, the multi-hour, many-step tasks where models usually lose the plot.
- 1M token context with text, image, and video input.
- Two launch variants: K3 Max for chat and agent tasks, and K3 Swarm Max for large-scale parallel processing, an evolution of the agent-swarm design Kimi has been building since K2.6.
What the benchmarks say
The launch numbers, with the usual caveat that several are self-reported until the weights land and independent replication catches up:
- 93.5% on GPQA Diamond, the strongest open-weight result published on that benchmark at release.
- 88.3% on Terminal-Bench 2.1, the benchmark closest to what an agent actually does inside an OutcomeDev sandbox: real work in a real shell.
- 91.2% on BrowseComp, the best published score at release, plus 84.2% on MCP Atlas and 56.0% on Humanity's Last Exam with tools. The pattern is consistent: K3's edge is agentic work, not trivia.
- #1 on the Frontend Code arena, ahead of every model from every lab, including Claude Fable 5. For UI-heavy builds, that is a remarkable data point.
- On the Artificial Analysis Intelligence Index v4.1, K3 scores 57.1, which slots it as roughly the third-strongest model family in the world, behind Anthropic's Fable 5 tier at 59.9 and OpenAI's GPT-5.6 Sol Max at 58.9, and ahead of essentially everything else.
Read those together and the picture is clear: the absolute frontier still belongs to Anthropic, but K3 sits close enough behind it that for many workloads, especially frontend-heavy ones, the difference will not be what you notice.
Why "competitive" is the right word, and where the catch is
Two things make K3 more interesting than its raw scores.
First, the weights. By July 27 you will be able to run the third-best model family in the world on hardware you control, under a permissive license. No closed frontier model offers that. For teams with data-residency requirements, or anyone hedging against API pricing changes, that matters more than a benchmark point or two.
Second, the price signals confidence. At $3 per million input tokens and $15 per million output, K3 costs roughly what Claude Sonnet costs, and several times what Kimi K2 did. Moonshot is saying this model competes on quality, not on discount. Whether the market agrees will show up in usage numbers, not press releases.
The honest catch: K3 currently ships with a single reasoning level and thinks hard by default. Early testing shows it spending north of ten thousand reasoning tokens on even simple tasks, and at $15 per million output tokens, that adds up. For quick, shallow tasks, a lighter model is still the economical pick. K3 earns its cost on the long, complex work its architecture targets.
Running K3 on OutcomeDev
K3 is live in the model selector today under the Kimi provider. OutcomeDev is strictly bring-your-own-key: add your Moonshot API key once in settings, select K3, and every sandbox, preview, and pull request works exactly as it does with every other model on the platform. Your prompts and code go to the provider you chose, on your key, and nowhere else.
If you want a feel for where K3 shines, give it something with a real horizon: a full returns-and-refunds desk, a scheduling tool with calendar integration, a frontend-heavy dashboard. Watch the agent narrate its way through, then check the arena results against your own experience.
The model multiverse keeps getting deeper. The best way to find your model is to run the same outcome through two of them and diff the results, and OutcomeDev is the easiest place in the world to do exactly that.