# Kimi K3 weights are live: what the 1.56 TB release actually opens

> Moonshot has published Kimi K3's full 2.8-trillion-parameter weights, license and technical report. The release is real, but serving it calls for data-center-scale hardware and the benchmark sheet remains vendor-reported.

_Source: Moonshot official release and public model weights · 2026-07-17 · 8 min read · Verified against primary sources_

Canonical: https://iyu.app/e/kimi-k3-open-3t-model

## The 60-second version

Moonshot has released Kimi K3 as a real, public and ungated open-weight model: 2.8T total parameters, 104B active per token, 1M context and about 1.56 TB of files.

**Key points**

- The Hugging Face repository contains 96 safetensors shards plus configuration and license files.
- K3 uses 16 of 896 experts per token, with MXFP4 weights and MXFP8 activations.
- Moonshot recommends supernodes with at least 64 accelerators, putting self-hosting beyond normal workstations.
- The custom license broadly permits modification and commercial use but adds thresholds for large MaaS operators and very large products.
- The weights are verified as public; benchmark scores remain vendor-reported until independently reproduced.

**Verdict.** This is a substantial open-weight release, not a repository placeholder. Its practical audience is researchers and large inference operators; most developers should test the hosted API first.

## Full explainer

> **✓** Update, July 28: the promised weights have arrived. Moonshot's official Hugging Face repository is public and ungated, with 96 safetensors shards, configuration files and a license. That verifies the release itself. Benchmark scores remain Moonshot's own measurements unless an independent evaluator reproduces them.


### What changed — The weights are no longer a promise

Kimi K3 launched as a claim about a future open release. It is now an inspectable artifact. The public repository contains the full model files, while GitHub carries the technical report and release documentation. The model download is about 1.56 TB in decimal units, or 1.42 TiB.

- **2.8T** — Total MoE parameters
- **104B** — Parameters active per token
- **1.56 TB** — Public model files
- **1M** — Token context window


### How it works — A huge model with a sparse active path

K3 is a Mixture-of-Experts model with 896 experts. Sixteen experts are selected for each token, plus two shared experts, so all 2.8 trillion parameters do not run on every step. Moonshot reports 104 billion active parameters per token. Its 93 layers combine 69 Kimi Delta Attention layers with 24 gated MLA layers, and a 401-million-parameter MoonViT-V2 encoder adds native image understanding.

- **Architecture:** Sparse Mixture-of-Experts; 896 experts, 16 selected per token and 2 shared experts.
- **Attention:** 69 Kimi Delta Attention layers plus 24 gated MLA layers.
- **Precision:** MXFP4 weights and MXFP8 activations, trained with quantization awareness.
- **Modalities:** Text and images in the released model; the product blog also demonstrates video-oriented workflows.
- **Context:** 1,048,576 tokens.


### Can you run it? — Open does not mean laptop-sized

The files can be downloaded without requesting access, but local deployment is a data-center project. Moonshot recommends supernodes with 64 or more accelerators. Even before runtime overhead, the 1.56 TB repository is far beyond ordinary workstation memory and storage budgets. The practical route for most developers is the hosted API or a specialist inference provider.

> **⚡** The useful distinction: K3 is open-weight, not easy-to-self-host. Researchers can inspect, modify and fine-tune it; most teams will still rent the compute.


### License — Broad rights with commercial thresholds

The custom Kimi K3 License permits use, copying, modification, redistribution, sublicensing, sale, deployment and fine-tuning. It is not Apache 2.0 or MIT. A Model-as-a-Service business whose group revenue exceeds $20 million across any consecutive 12 months must reach a separate agreement with Moonshot before commercial use. Products above 100 million monthly active users or $20 million in monthly revenue must prominently display “Kimi K3.” Internal use is exempt from those two conditions.

- **Research and modification:** Allowed, including fine-tuning and derivative works.
- **Commercial products:** Generally allowed, subject to the license thresholds and attribution conditions.
- **Large MaaS operators:** Separate agreement required above $20M aggregate revenue over any consecutive 12 months.
- **Very large products:** Prominent “Kimi K3” display required above 100M MAU or $20M monthly revenue.


### Performance claims — The scoreboard still needs outside testing

Moonshot reports 88.3 on Terminal-Bench 2.1, 67.5 on DeepSWE and 93.5 on GPQA Diamond, alongside broad multimodal and knowledge-work results. Those numbers make K3 look competitive with leading closed systems on several tasks, but the release does not turn vendor measurements into independent evidence.

> **⚑ Caveat:** Treat the benchmark table as a vendor claim. Harnesses, reasoning budgets, fallback behavior and hardware differ across compared models. The newly public weights make independent evaluation possible; they do not provide it automatically.


### Access and pricing — API access is the realistic starting point

The official API model name is `kimi-k3`. Moonshot lists cached input at $0.30 per million tokens, uncached input at $3.00 and output at $15.00. It says coding workloads exceed a 90% cache-hit rate, another figure teams should verify against their own prompt patterns.

> K3 has crossed the important line from announcement to downloadable weights. The next test is not whether the files exist, but whether independent users can reproduce the capability at a cost that makes sense.


### Bottom line — A real open-weight milestone with a steep compute bill

Kimi K3 is now a genuine open-weight release: public files, a technical report and a license that grants meaningful rights. Its scale also defines the catch. The model expands what researchers and infrastructure providers can inspect and build on, while ordinary developers will mostly experience it through APIs. Download the weights for research or deployment planning; choose the API for practical evaluation; and wait for independent benchmarks before treating the frontier claims as settled.


## Primary sources

- [Moonshot AI — Kimi K3 official repository](https://github.com/MoonshotAI/Kimi-K3)
- [Hugging Face — moonshotai/Kimi-K3 weights](https://huggingface.co/moonshotai/Kimi-K3)
- [Moonshot AI — Kimi K3 technical blog](https://www.kimi.com/blog/kimi-k3)
- [Kimi K3 License](https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE)

---
_Published by iyu (https://iyu.app) — the day's AI news, checked against primary sources and rewritten in plain language. Free to quote with attribution and a link to the canonical URL._