Kimi K3: Why Moonshot Gave Away a 2.8-Trillion-Param Model

Silicon Valley spent a chunk of July on edge about a model it can't even fully inspect yet. Moonshot AI launched Kimi K3 through its API in mid-July 2026, claiming it beat several current-generation US flagship models on public benchmarks while trailing only the very newest frontier releases from OpenAI and Anthropic. Then, about a week later, Moonshot did something a well-funded, revenue-generating US lab almost never does: it published the actual model weights, letting anyone download and run K3 on their own infrastructure.

What actually shipped

Kimi K3 is a mixture-of-experts model with 2.8 trillion total parameters, of which roughly 104 billion activate for any given token — a design that keeps inference cost far below what a dense model that size would cost to run. It carries a 1-million-token context window with native text and image input, and Moonshot's own hosted API prices it at $3 per million input tokens and $15 per million output tokens, flat across the full context length with no long-context surcharge.

On performance, the honest read is: strong, not universally best. Moonshot itself frames K3 as outperforming most current models except the latest from OpenAI and Anthropic — a claim independently echoed by outlets covering the release rather than an internal marketing number.

The open-weight play isn't generosity

US export controls dating to 2022 have limited Chinese labs' access to the advanced chips frontier training runs need. Giving weights away is a direct response: it spreads a model's footprint across compute the lab doesn't own, builds an ecosystem, and puts pressure on the closed, API-only pricing of Western frontier labs — without requiring Moonshot to out-buy anyone on GPUs. It's working at the margins: Chinese open-weight models accounted for 17.1% of global AI model downloads in the year to August 2025, edging past the US's 15.86% for the first time on record. Alibaba's Qwen (Apache 2.0) and DeepSeek's MIT-licensed releases were already pulling in that direction before K3 arrived; Alibaba answered days later with its own Qwen 3.8 claims.

Read the license before you build on it

'Open weight' is doing a lot of work in Kimi K3's case. Moonshot published the weights under a bespoke Kimi K3 License, not a permissive standard like MIT or Apache 2.0. Two clauses matter if you're evaluating it for anything beyond a prototype: run it as a model-as-a-service offering — giving third parties inference or fine-tuning access — and once your revenue (plus affiliates') crosses $20M over any trailing 12 months, you're required to sign a separate commercial agreement with Moonshot. Separately, any product or service with more than 100 million monthly active users, or more than $20M in monthly revenue, must display 'Kimi K3' branding in its interface. Neither clause blocks experimentation or internal tooling, but they mean the free-and-open framing has a ceiling — and that ceiling is exactly where a growing product would hit it.

Why builders should care

For teams evaluating LLMs for trading tools, agentic pipelines, or any AI-assisted product, K3's release is a useful data point in two directions. First, self-hostable frontier-adjacent models at this parameter-efficiency level (104B active of 2.8T total) genuinely change the self-hosting cost-benefit math versus renting inference from a closed API. Second, model choice is no longer just a benchmark and price comparison — it's a license review. A model that's free today can carry commercial terms that only bite once you've built something people actually use, which is precisely the point in a product's life when switching providers is most expensive.

Chinese open-weight models accounted for 17.1% of global AI model downloads in the year to August 2025, narrowly surpassing the US's 15.86% for the first time.

— Reporting cited by The Verge