---
title: "Kimi K3: Moonshot pushes the open-weight frontier toward 3 trillion parameters"
description: "Moonshot AI unveils Kimi K3, which it calls 'the world's first open 3T-class model': a Mixture of Experts with 2.8 trillion parameters (16 of 896 experts active), a 1-million-token context window, long-horizon agentic and coding abilities and native vision. Available today via API and apps; the open weights are expected by July 27, 2026. An open frontier closing on the best proprietary models, with the limits the maker itself discloses."
url: "https://www.noze.it/en/insights/kimi-k3-moonshot-open-weight-frontier/"
lang: "en"
alternate: "https://www.noze.it/insights/kimi-k3-moonshot-frontiera-pesi-aperti/"
published: "2026-07-17"
updated: "2026-07-28"
author: "noze S.r.l."
keywords: "AI, Open Source, Kimi K3, Moonshot AI, Open Weights, MoE, Agentic Coding, Long Context, Vision, On-Premise, AI Sovereignty, LLM"
---

# Kimi K3: Moonshot pushes the open-weight frontier toward 3 trillion parameters

Moonshot AI unveils Kimi K3, which it calls 'the world's first open 3T-class model': a Mixture of Experts with 2.8 trillion parameters (16 of 896 experts active), a 1-million-token context window, long-horizon agentic and coding abilities and native vision. Available today via API and apps; the open weights are expected by July 27, 2026. An open frontier closing on the best proprietary models, with the limits the maker itself discloses.

17 July 2026Updated 28 July 2026

![Bar chart of Kimi K3's coding benchmarks compared with GPT-5.6 Sol, Fable 5, Opus-4.8, GPT-5.5 and GLM-5.2 across DeepSWE, Terminal Bench 2.1, FrontierSWE, Program Bench, Kimi Code Bench 2.0 and SWE Marathon](https://www.noze.it/images/insights/kimi-k3-coding-bench.png)

The six coding benchmarks reported by Moonshot: Kimi K3 (in blue) leads on Program Bench and SWE Marathon, sits half a point behind GPT-5.6 Sol on Terminal Bench 2.1 and falls further back on DeepSWE. Image: Moonshot AI (kimi.com).

## What Kimi K3 is

**Moonshot AI** has unveiled **Kimi K3**, which it calls _“the world’s first open 3T-class model.”_ It is a **Mixture of Experts** with **2.8 trillion** total parameters, of which it activates only **16 out of 896 experts** per token, with a context window of **1 million tokens** and **native vision**. The positioning is clear: frontier intelligence for long-horizon coding, knowledge work and reasoning.

> **Note.** The benchmarks, figures and prices here come from Moonshot AI’s announcement, not from independent testing by noze; the benchmark results are the vendor’s own claims. As of July 17, 2026 the model is usable via API and apps, but **the open weights have not been released yet**: where that matters, we say so.

## Under the hood

The architecture is all about efficiency at this scale. Moonshot claims roughly a **2.5×** improvement in scaling efficiency over Kimi K2, thanks to several choices:

- **Kimi Delta Attention (KDA)**, an attention form designed to scale efficiently over long contexts.
- **Attention Residuals (AttnRes)**, to selectively retrieve representations across the model’s depth.
- A **Stable LatentMoE** framework and **Gated MLA** (multi-head latent attention).
- **Quantization-aware** training, with weights in **MXFP4** and activations in **MXFP8**: quantization is part of training, not a fallback bolted on afterwards.

The theme is the one we have been tracking for a while: huge models made runnable through compressed attention and aggressive quantization, as with [DwarfStar 4](https://www.noze.it/en/insights/dwarfstar-4/) and [colibri](https://www.noze.it/en/insights/colibri-glm-5-2-local/). Here the scale is another matter, though: 2.8 trillion parameters are a lot even just to store.

## What it can do

Kimi K3 is built for **long-horizon agentic use**: sustained engineering sessions with little oversight, repository navigation and terminal orchestration. The examples Moonshot shows are ambitious: a GPU compiler (**MiniTriton**) written from scratch, autonomous **48-hour** runs on chip design, game development with _vision-in-the-loop_ iteration, and on the research side a 42-year analysis of the ASIC industry with over **2,800 web searches** and **11,000 pages** read, or a gravitational-wave analysis with more than **20 subagents** in parallel.

On benchmarks (at maximum thinking effort, figures reported by Moonshot) the numbers are high: **Terminal-Bench 2.1 at 88.3**, **GPQA-Diamond at 93.5**, **DeepSWE at 67.5**, **MMMU-Pro (vision) at 81.6**, **CharXiv at 84.8**. Moonshot is candid, though: _“overall performance still trails the most powerful proprietary models,”_ namely [Claude Fable 5](https://www.noze.it/en/insights/fable-5-shutdown-us-government-ai-sovereignty/) and [GPT-5.6 Sol](https://www.noze.it/en/insights/gpt-5-6-sol/). On some tests, like Terminal-Bench, K3 closes on or beats them; on others it stays a step behind.

From today it is available via web (Kimi.com), a desktop app (**Kimi Work**, Windows and Apple Silicon Macs), the terminal (**Kimi Code**) and the **API** (platform.kimi.ai). API prices: **$0.30/Mtok** for cache-hit input, **$3.00/Mtok** for cache-miss input, **$15.00/Mtok** for output, with a claimed cache-hit rate **above 90%** in coding workloads. At launch the model defaults to maximum thinking effort; low and high modes will come later.

## The limits it discloses

Credit to Moonshot for listing the flaws. The first is technical and interesting: **sensitivity to thinking history**. If the harness does not pass back to the model all the _thinking_ content produced in previous turns, _“generation quality may become highly unstable.”_ It is a reminder that, with agents, the environment around the model matters as much as the model itself, which is exactly the point of [harness engineering](https://www.noze.it/en/insights/harness-engineering/). The other two disclosed limits: **excessive proactiveness**, with unexpected autonomous decisions, and a **gap in user experience** compared with Fable 5 and GPT-5.6 Sol.

## The catch: the weights land on July 27

And here is the point that matters most to us. Kimi K3 is announced as an **open-weight** model, but as of July 17, 2026 the weights are **not downloadable yet**: Moonshot states that _“the full model weights will be released by July 27, 2026.”_ Until then K3 is, in practice, a **cloud API** like any other, with the same questions about what leaves your perimeter when you hand it code and data, as we showed for [AI coding CLIs](https://www.noze.it/en/insights/ai-coding-cli-telemetry/).

The sovereignty payoff arrives when the weights are truly in your hands and you run them yourself. But two concrete issues remain: a 2.8-trillion-parameter MoE is enormous to run in-house, and it will need the kind of quantization and expert-streaming that engines like [colibri](https://www.noze.it/en/insights/colibri-glm-5-2-local/) provide to fit on reasonable hardware; and a release date, as always, is a promise until it materialises.

**Update, 28 July: the weights are out.** Moonshot published Kimi K3 on Hugging Face on **26 July**, a day ahead of the deadline it had set itself. The weights live at [huggingface.co/moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3) in **safetensors** format, and the [MoonshotAI/Kimi-K3](https://github.com/MoonshotAI/Kimi-K3) repository hosts the technical report. There was no separate release post: the announcement is still [the one from 16 July](https://www.kimi.com/blog/kimi-k3), and the weights simply appeared.

The sizes confirm the problem we flagged: roughly **594 GB** in the native MXFP4 format, up to **1.4 TB** depending on quantization. At that size, downloading and running it takes planned storage and hardware.

One clarification we care about, because it is exactly the distinction we keep insisting on. The licence is titled the **Kimi K3 License** and is an MIT with two added clauses: anyone operating a language-model service with aggregate revenue **above USD 20 million** over twelve months must agree commercial use separately, and products above **100 million monthly active users** or USD 20 million in monthly revenue must display “Kimi K3” prominently in the user interface. Internal use and access through Moonshot’s official products are exempt from both. For the vast majority of organisations these thresholds are irrelevant, and the weights remain downloadable and runnable without asking permission. It is still an **open-weight licence with commercial conditions**, though, not an OSI-approved open source licence: the two should not be conflated.

## What we think

Kimi K3 confirms a trajectory we have been following piece by piece: the **open-weight** frontier keeps closing on the best proprietary models, after [GLM 5.2](https://www.noze.it/en/insights/glm-5-2-open-weight-ai-sovereignty/), [Ornith 1.0](https://www.noze.it/en/insights/ornith-1-0-self-scaffolding/) and [Inkling](https://www.noze.it/en/insights/inkling-thinking-machines-open-weights/) just two days ago. An open 3T-class model, competitive on coding and agents, is important news for those who want to own intelligence rather than rent it.

With two caveats we always repeat. First: a model is open when the weights are downloadable, not when it is announced; until July 27, for us, it stays an API to be judged as such. Second: owning the model is half the job, the other half is **governing** it, with an audit trail and policies over any model, local or remote, as [Admina](https://www.noze.it/en/open-source/admina/) does. That is the sense of [Open Intelligence, Secure Governance](https://www.noze.it/en/insights/open-intelligence-secure-governance-agentic/), and the direction we work in: [on-premise AI](https://www.noze.it/en/artificial-intelligence/), sovereign and governed, on [Linux infrastructure](https://www.noze.it/en/linux-services-systems/) that stays yours. If the weights really arrive on July 27, it will be another brick in that operational floor.

## Sources

- [Kimi K3 (Moonshot AI official blog)](https://www.kimi.com/blog/kimi-k3)
- [Kimi K3 weights on Hugging Face](https://huggingface.co/moonshotai/Kimi-K3)
- [MoonshotAI/Kimi-K3 on GitHub (technical report)](https://github.com/MoonshotAI/Kimi-K3)
- [Inkling: Thinking Machines’ open-weights model (noze)](https://www.noze.it/en/insights/inkling-thinking-machines-open-weights/)
- [colibri: GLM 5.2 locally, streaming the experts from disk (noze)](https://www.noze.it/en/insights/colibri-glm-5-2-local/)
