Quick Take
- Moonshot AI released Kimi K3 open weights on July 27, 2026, for free public download.
- Kimi K3 runs 2.8 trillion parameters, the largest open-weight AI model in the world.
- Indian builders can self-host Kimi K3 so user data stays inside India, not on foreign servers.
In This Article
Kimi K3, the largest open-weight AI model in the world, is now free to download after Moonshot AI released its weights on July 27, 2026, from Beijing. The move lets developers everywhere run the model on their own machines at no licence cost.
The Beijing company, valued above $20 Bn (Rs 1,92,400 Cr) after a $2 Bn (Rs 19,240 Cr) raise in May 2026, built Kimi K3 with 2.8 trillion parameters and a one-million-token context window. Founder Yang Zhilin says openness will win users faster than closed United States systems. Moonshot published the Kimi K3 weights on Hugging Face, ten days after the model first went live through an app and paid service.
StartupFeed Insight
The headline is the 2.8 trillion parameter count, but the real signal for India sits elsewhere. The hosted Kimi K3 service is not the cheap option, so the win for Indian builders is self-hosting: prompts and user data stay inside India. StartupFeed expects data control, not price, to drive adoption across regulated sectors like banking and health-tech. Watch IndiaAI Mission teams and mid-size IT firms closely. Our call: at least one major Indian IT services firm or IndiaAI-backed group will ship a Kimi K3 fine-tune or an Indian-language variant before December 31, 2026. By Avinash.
Kimi K3 by the numbers: what was released?
Kimi K3 is a 2.8 trillion parameter open-weight LLM (Large Language Model) released by Moonshot AI on July 27, 2026. The company shared the model weights, the settings that steer an AI toward its answers, so anyone can download, tweak and host it.
| Metric | Detail | Notes |
|---|---|---|
| Model | Kimi K3 (open-weight LLM) | Weights free on Hugging Face |
| Parameters | 2.8 trillion | Largest open-weight model to date |
| Context window | 1 million tokens | Fits large codebases in one query |
| Weights release | July 27, 2026 | Ten days after the app and API debut |
| Developer | Moonshot AI, Beijing | Founded 2023 |
| Backing | $2 Bn (Rs 19,240 Cr) raise, May 2026 | Valued above $20 Bn (Rs 1,92,400 Cr) |
The full download runs about 594 GB in its native format, so this is a heavy file, not a quick install. One point matters for buyers: open weight is not the same as open source, and some analysts noted no clear licence file at launch.
About Moonshot AI
Moonshot AI is a Beijing artificial intelligence company founded in 2023 by Yang Zhilin, building the Kimi family of models and the Kimi assistant. It uses a mixture-of-experts (MoE) design, where only part of the model runs on each query to cut compute cost. Backers include Alibaba, Tencent and the investment arm of food-delivery firm Meituan. The company raised $2 Bn (Rs 19,240 Cr) in May 2026 at a valuation above $20 Bn.
Why did Kimi K3 rattle global AI stocks?
Kimi K3 rattled global AI stocks because it showed a Chinese lab can build a frontier-class model at lower cost, which questions the case for heavy AI spending. The PHLX Semiconductor Index fell about 10% in the launch week, its steepest weekly drop since April 2025, according to Wall Street Journal reports. Nvidia fell more than 2% as traders repriced chip demand. Global chip stocks have shed close to $3.3 Tn (about Rs 317 Lakh Cr) since late June, a broader slide that Kimi K3 then sharpened, market reports said.
Openness and wider availability, not a closed system, will win more users over time, according to Yang Zhilin, Founder, Moonshot AI.
The reaction mirrored the DeepSeek moment of January 2025, when another Chinese model matched United States rivals at a fraction of the price. Investors are not only selling chips. They are testing the belief that top AI needs an ever-growing, United States-controlled compute build-out.
What does Kimi K3 mean for Indian developers?
For Indian developers, Kimi K3 matters most as a model they can self-host, keeping data inside India rather than routing it through a China-linked API (Application Programming Interface). The hosted service is not cheap for India, at about Rs 258 per one million input tokens and Rs 1,290 per one million output tokens, as reported. So the story is capability you can eventually run yourself, not a lower bill.
Native vision is the other draw. Kimi K3 can read screenshots and layouts and turn them into code, which helps Indian product teams skip a separate vision model. This lands as the IndiaAI Mission pushes a home-grown foundational model, with IT Minister Ashwini Vaishnaw targeting an Indian model within months. Before routing sensitive work through any China-linked service, Indian firms should still weigh China’s National Intelligence Law and the missing licence detail.
How does Kimi K3 compare to top AI models?
Kimi K3 ranks third on the independent Artificial Analysis Intelligence Index, behind Anthropic’s Claude Fable 5 and OpenAI‘s GPT-5.6 Sol. Moonshot says its model performs competitively with Fable 5 and outperforms Anthropic’s Opus 4.8 and OpenAI’s GPT-5.6 Sol on several tests, a claim developers can now check for themselves.
| Model | Type | Standing |
|---|---|---|
| Kimi K3 (Moonshot) | Open-weight, 2.8T | Ranked third, self-hostable |
| Claude Fable 5 (Anthropic) | Proprietary API | Top of the index |
| GPT-5.6 Sol (OpenAI) | Proprietary API | Second on the index |
| DeepSeek V4 | Open-weight | Ranks below Kimi K3 |
One caution sits next to the cost win: an independent test flagged a high hallucination rate near 51% for the model, so accuracy checks are wise. What sets Kimi K3 apart is scale plus native vision in a model you can host yourself, though it needs far heavier hardware than most open models.
What’s Next
Moonshot is expected to publish a technical report on how Kimi K3 was trained and tested. The company has also floated a Hong Kong IPO (Initial Public Offering) within six months, per shareholder documents. The real test now is distribution: whether major cloud platforms host the model and whether enterprises trust it with live data. Which side will Indian firms land on?
Frequently Asked Questions
Disclaimer: This article is for informational purposes only and does not constitute investment advice. StartupFeed and its authors are not SEBI-registered investment advisors. The analysis above is based on publicly available information and should not be the sole basis for any investment decision. Please consult a SEBI-registered financial advisor before making investment decisions.
Have a tip? Write to us at editorial@startupfeed.in.
