Moonshot AI has released Kimi K3, a new model already being compared to Fable 5 and GPT-5.6. It has quickly been dubbed a "free competitor" to the flagship models, though that label comes with caveats: only a stripped-down tier and, unsurprisingly, open weights for self-hosting are genuinely free. Cloud access and the multi-agent mode are paid.
Under the hood sits roughly 3 trillion parameters and a context window of up to one million tokens. In practice, this means the model can hold an entire book, a large code repository, or a stack of lengthy documents in its "working memory" at once — exactly the kind of scenario where ordinary models tend to lose the thread and start mixing things up.
K3 can be accessed via API, and for those who just want to jump in and try it, there is a Kimi K3: Max version with a chat interface and dedicated coding modes.
One particularly interesting piece is K3 Swarm — a setup involving several agents that split a large task among themselves: one digs through the problem, another verifies the findings, and a third assembles the final result. This is exactly the direction the industry is heading: not "the single smartest chatbot," but a coordinated team of specialized agents handling a specific job.
How much it costs
Kimi does offer a free tier, but it's limited: capped usage and no access to Thinking mode. Serious work requires a paid plan:
Moderato — $19/month: the entry-level paid tier with higher quotas.
Allegretto — $39/month: adds Instant and Thinking modes.
Allegro — $99/month: full access to Agent Swarm, the multi-agent mode.
Vivace — $199/month: maximum usage limits.
Via API, pricing is per token — early estimates suggest it runs roughly half the cost per token compared to GPT-5.6.
There's also a separate route: download the open weights and run the model on your own infrastructure. In that case, "free" is genuinely free — the only real cost is your own GPU hardware.
What matters here isn't just the model's size, but the fact that a serious frontier-level model is once again being released with open weights. While others wait for access to the next closed-source flagship, K3 can be downloaded and deployed independently right now — complete with a million-token context window and a multi-agent mode built in.
If you're deciding which model to build your workflows on and don't want to overpay for a flagship-tier system, Kimi K3 is at least worth experimenting with. Between its open-weight availability, generous context window, and built-in multi-agent capabilities, it's a strong signal that the gap between open and closed frontier models continues to narrow.