# Kimi K3 Weights Expected Today: 2.8T Free Download

Moonshot AI is expected to release the 2.8T-parameter Kimi K3 weights today under a Modified MIT license. A Hugging Face countdown was still ticking. The largest open-weight release in history is hours away.

By Neomanex. Updated 2026-10-07.

Moonshot AI is expected to release the full [Kimi K3](/models/kimi-k3/) weights for public download on July 27, 2026 — making good on the promise it made at the model's July 16 launch. If the weights land, it's the single most significant open-weight event of 2026 — a frontier-adjacent model handed to the world with a permissive Modified MIT license, racing a US sanctions clock that could close the window. Our verdict on K3 stays conditional. As of this writing, a countdown clock on the [official Hugging Face repo](https://huggingface.co/moonshotai/Kimi-K3) was still ticking, with 1,560 people signed up for release notifications. The 2.8-trillion-parameter Mixture-of-Experts model will be downloadable under a Modified MIT license. At roughly 1.4TB (the MXFP4-quantized weight format that Moonshot trained the model with), it's expected to be the largest open-weight release ever by a margin of approximately 3x over the next largest downloadable model.

The release, if it lands, represents the single most significant open-weight event of 2026: a frontier-adjacent model that leads the Arena WebDev leaderboard being handed to the world with a permissive license. The timing is inextricable from geopolitics — Moonshot is shipping before potential US regulatory intervention can close the window.

## What happened

Kimi K3 has been accessible through Moonshot's API and hosted platforms since July 16. Its pricing — $3/$15 per million tokens with 90%+ cache-hit rates on coding workloads — and its #1 [Arena WebDev](https://arena.ai/leaderboard/code) score (1682) were already public. What Moonshot promised for July 27: downloadable weights.

The Modified MIT license means any developer will be able to download, modify, fine-tune, and self-host K3. The ~1.4TB download uses a Mixture-of-Experts architecture — 896 total experts, 16 activated per token — which keeps inference costs manageable at API scale. For self-hosting, the Hugging Face community analysis estimates 8-16 nodes of 8× H100/B200 GPUs (64-128 accelerators) are needed to load the full model.

## Why it matters

**Sovereignty.** Self-hosting eliminates API dependency, rate limits, and data egress. For enterprises running sensitive workloads, the difference between using Kimi K3 through an endpoint and running it on your own infrastructure is the difference between a vendor relationship and an architectural choice.

**The policy clock is running.** The White House has already accused Moonshot of industrial-scale IP theft via distillation of Anthropic's [Fable 5](/models/claude-fable-5/) (July 22). Treasury Secretary Bessent threatened sanctions and Entity List designations. Regulatory scrutiny is intensifying — see our [coverage of the White House accusations against Moonshot](/news/white-house-accuses-moonshot-fable-distillation-kimi/). Once weights are downloaded, they cannot be recalled — and Moonshot appears to be racing the regulatory window.

**The coalition grew in lockstep.** The US open-weight coalition letter hit 50 signatories this week — a direct pre-response to the expected crackdown on Chinese open-weight releases (India Today, 2026).

**Dean Ball called it "AI communism."** OpenAI's Head of Strategic Futures described the open-weight model trajectory as "full AI communism," where AI becomes a state-provided public good, calling it a "dystopian hellscape" (Ball, via TechCrunch, 2026). The framing captures US anxiety: a frontier-capable Chinese model matching second-tier American performance, released for free, undercutting the revenue models of every closed-model lab charging premium API prices.

For our directory: K3's technical strengths and weaknesses haven't changed. It leads Arena WebDev at 1682 (as of July 26). It still trails Fable 5 and [GPT-5.6 Sol](/models/gpt-5-6-sol/) on general reasoning. The White House investigation and sanctions threat remain material risks — our verdict stays **conditional**. But if the weights drop today, the calculus shifts: K3 moves from "not yet downloadable" to "available for self-hosting evaluation." The Modified MIT license would put K3 in the permissively-licensed open-weight tier alongside models like GLM-5.2 (MIT) and Hy3 (Apache 2.0).

## What changes for you

| If you... | Then... |
|-----------|---------|
| Have large-scale GPU infrastructure (est. 64+ accelerators) | Prepare to download K3 from Hugging Face when the countdown reaches zero — the Modified MIT license permits evaluation and self-hosting |
| Run a cloud inference platform | The Modified MIT license makes hosted K3 straightforward — expect providers to add it within weeks of the weight drop |
| Use K3 through the API today | No immediate change — the API remains available. Self-hosting will become an option once weights land |
| Don't have the hardware for ~1.4TB | Watch which cloud providers add hosted K3 inference, or continue using Moonshot's API at $3/$15 per million tokens |

The infrastructure reality is blunt: this is not a laptop model. The ~1.4TB download itself is the first gate. But for teams already running large open-weight infrastructure, K3's weights would be a real deployment option rather than a promise — once the countdown reaches zero.

## FAQ

**Are the weights live right now?**

As of this writing, the [Hugging Face countdown clock](https://huggingface.co/moonshotai/Kimi-K3) had not yet reached zero. The page showed "Upcoming release" with a timer counting down to July 27, 2026. The weights are expected today — they have not been released at the time of publication.

**Can I actually download and run Kimi K3 once the weights drop?**

Yes — once live. The Modified MIT license permits download, modification, fine-tuning, and self-hosting. You'll need substantial GPU infrastructure (estimated 64+ accelerators) and sufficient storage for the ~1.4TB download — the Hugging Face community analysis estimates 8-16 nodes of 8× H100/B200 GPUs to load the model.

**Does this change the White House investigation?**

No. The IP-theft accusation and sanctions threat are independent of weight availability. If anything, the release may accelerate the regulatory response — weights that have been downloaded globally are harder to restrict than weights still behind an API gate.

**How does K3 compare to other open-weight models?**

K3's 2.8T parameters would dwarf the next largest open-weight model by roughly 3x. GLM-5.2 (MIT license) and Hy3 (Apache 2.0) are smaller but have more permissive licensing and lower self-hosting requirements. Poolside's Laguna S 2.1 is a strong Western alternative at 118B MoE (8B active). For coding workloads specifically, K3's Arena WebDev lead (1682) is the strongest signal among open-weight models.

Canonical: https://neomanex.com/news/kimi-k3-open-weights-live