# Kimi K3 Open Weights Now Live on Hugging Face

Moonshot AI released the full Kimi K3 open weights (2.8T parameters) on Hugging Face on July 27 under the Kimi K3 License, a custom license with commercial thresholds. Self-hosting eliminates China data-sovereignty risk for regulated teams.

By Neomanex. Updated 2026-10-07.

Moonshot AI delivered on its promise. On July 27, 2026 at midnight UTC (evening of July 26 ET), the company published the full open weights for **Kimi K3** on its Hugging Face organization under the Kimi K3 License — a custom license with commercial use thresholds ($20M trailing-12-month revenue, 100M monthly active users). The 2.8-trillion-parameter Mixture-of-Experts model — the largest open-weight release ever — is now self-hostable, downloadable, and inspectable by anyone with the hardware to run it.

## What happened

Since July 16, K3 has been available through Moonshot's API at $3/$15 per million tokens (input/output). But the weights were locked down — developers couldn't inspect, fine-tune, or self-host the model.

The full weights (~594 GB MXFP4 format) are now downloadable from [Moonshot's Hugging Face organization](https://huggingface.co/moonshotai/Kimi-K3). vLLM support shipped day-0 with Kimi Delta Attention (KDA) prefill cache integration. [Together AI](https://www.together.ai/models/kimi-k3) and [Modal](https://modal.com/blog/kimi-k3-by-moonshot-now-available-on-modal) both announced same-day hosted access, timed to the weights drop. [OpenRouter](https://openrouter.ai/moonshotai/kimi-k3) followed shortly after.

| Spec | Value |
|------|-------|
| Total parameters | 2.8 trillion (MoE, 16 of 896 experts active per token) |
| Context window | 1,048,576 tokens (1M) |
| Weight format | MXFP4 (~594 GB download) |
| License | Kimi K3 License (custom; $20M revenue / 100M MAU commercial thresholds) |
| API pricing (since Jul 16) | $3.00 input / $15.00 output per 1M tokens |
| Hosted day-0 | Together AI, Modal, OpenRouter |
| Self-host minimum | 64+ accelerators, ~1.4TB GPU memory |
| Prod-ready at full context | Q4 2026 (expected) |

**License detail:** The Kimi K3 License is NOT a Modified MIT license — it is a custom license that requires a separate commercial agreement from Moonshot for Model-as-a-Service operators exceeding $20M in aggregate trailing-12-month revenue, and requires "Kimi K3" attribution on products exceeding 100M MAU or $20M monthly revenue. Verify the actual [LICENSE file](https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE) before any commercial deployment.

## Why it matters

Self-hosting eliminates the single biggest risk that kept regulated teams away from K3: **prompts no longer touch servers subject to China's National Intelligence Law**. For enterprises with data sovereignty requirements, healthcare organizations, and financial institutions, this transforms K3 from a compliance liability into a deployable asset.

The open-weights release also enables independent inspection for the first time. Outside researchers can now examine MoE routing, tokenizer structure, and training artifacts — particularly relevant given the [unresolved distillation allegations](/directory/models/kimi-k3) from White House OSTP Director Michael Kratsios on July 22.

**Hardware reality check:** self-hosting requires 64+ accelerators and approximately 1.4TB of GPU memory. Production-stable full-context (1M token) inference is expected Q4 2026. For most teams, hosted access through Together AI, Modal, or OpenRouter is the pragmatic path today.

## What changes for you

**If you have the hardware:** Download the weights from [Moonshot's Hugging Face organization](https://huggingface.co/moonshotai/Kimi-K3), pin a commit hash, and verify the [LICENSE file](https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE) directly before commercial deployment. vLLM ships with KDA support out of the box.

**If you don't:** [Together AI](https://www.together.ai/models/kimi-k3), [Modal](https://modal.com/blog/kimi-k3-by-moonshot-now-available-on-modal), and [OpenRouter](https://openrouter.ai/moonshotai/kimi-k3) all offer same-day hosted inference at standard per-token prices — no GPU provisioning required.

**Everyone:** Re-run your own eval suite. Independent benchmarks can now verify vendor-reported scores against the Arena WebDev #1 claim (1,679, preliminary) and other published results. The fine-tuning ecosystem will build out in the weeks following release.

**Sanctions watch:** The White House investigation remains active. OSTP Director Kratsios's July 22 distillation allegations, Treasury's sanctions threats, and BIS's formal investigation are unresolved. But once public, weights are effectively impossible to retract — and 50 companies signed a July 27 open-weight coalition letter opposing restrictions.

## FAQ

**Is Kimi K3 actually the largest open-weight model?**

Yes, among tracked 2026 releases. At 2.8T total parameters, K3 surpasses DeepSeek V4 Pro (1.6T) and MiMo V2.5 Pro (1.02T). However, with only 16 of 896 experts active per token, the real compute footprint per forward pass is a fraction of the headline number.

**Can I run it on a single GPU?**

No. The full model requires 64+ accelerators and ~1.4TB GPU memory minimum. Consumer GPUs cannot load it even quantized. Use hosted inference through [Together AI](https://www.together.ai/models/kimi-k3), [Modal](https://modal.com/blog/kimi-k3-by-moonshot-now-available-on-modal), or [OpenRouter](https://openrouter.ai/moonshotai/kimi-k3) unless you have enterprise-grade infrastructure.

**What's the legal risk given the White House investigation?**

The investigation is active and unresolved. Kratsios's distillation allegations and Treasury's sanctions threats create real uncertainty for US-based commercial deployment. The Kimi K3 License governs use of the weights, but a potential Entity List designation or sanctions action could affect US companies hosting or deploying the model. We rate K3 [Conditional](/directory/models/kimi-k3) specifically because this legal and operational risk is material.

**Is this a Modified MIT license?**

No. The Kimi K3 License is a custom license — not Modified MIT. It imposes two commercial thresholds absent from MIT: (1) a separate commercial agreement requirement for Model-as-a-Service operators exceeding $20M trailing-12-month revenue, and (2) mandatory "Kimi K3" attribution on products exceeding 100M MAU or $20M monthly revenue. Moonshot's earlier K2-line models used Modified MIT, but that did not carry over to K3.

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