Kimi K3: Moonshot AI Just Launched the World's Largest Open-Weight Model — And It Is Gunning for Claude and GPT

China's Moonshot AI just dropped a model that has the AI world paying attention.

Kimi K3 launched on July 16, 2026, with 2.8 trillion parameters, a one-million-token context window, native vision, and always-on reasoning. At that parameter count, it is the largest open-weight model ever released — and on independent benchmarks, it is already ranking fourth among all frontier models worldwide, trailing only Claude Fable 5 and GPT-5.6 Sol.

The full open weights — which anyone can download, inspect, and run — are scheduled to drop by July 27, 2026.

Here is everything you need to know.

What Is Kimi K3?

Kimi K3 is the flagship model from Moonshot AI, the Beijing-based startup behind the Kimi chatbot and the K2 model family. If you have been following Moonshot AI's Kimi K2 or have used Kimi K2 on OpenRouter, K3 is the next major step — a new-architecture model that consolidates and significantly extends everything the K2 line established.

K3 is a Mixture-of-Experts model with 2.8 trillion total parameters. Of those 2.8 trillion, only 16 out of 896 experts are active per token during inference. This design — large total capacity, focused active computation — is what lets the model achieve frontier-level performance without proportionally frontier-level inference costs.

Two variants shipped at launch. K3 Max is the general-purpose model for chat, knowledge work, and agentic tasks. K3 Swarm Max is designed for large-scale parallel processing — running many agent instances simultaneously across complex workflows.

Technical Specs

Spec	Detail
Total parameters	~2.8 trillion
Active parameters per token	~32 billion (16 of 896 experts)
Architecture	Mixture-of-Experts, Kimi Delta Attention
Context window	1 million tokens
Modalities	Text, image, video (native vision)
Reasoning	Always-on thinking mode
Open weights	July 27, 2026 (Modified MIT license)
Variants	K3 Max, K3 Swarm Max

The architecture is built on Kimi Delta Attention — a hybrid linear attention mechanism that Moonshot developed specifically for K3, combined with Attention Residuals. Moonshot says this architecture enables more efficient long-context processing compared to standard attention mechanisms, which is part of why a 1-million-token context window is practical at this scale.

Unlike the K2 series, which had separate standard and reasoning variants, K3 ships with reasoning always on. Moonshot calls this thinking mode, and the model works through problems by default rather than needing to be prompted into a reasoning mode.

Pricing

Token Type	Price
Cached input	$0.30 per million tokens
Standard input	$3.00 per million tokens
Output	$15.00 per million tokens

Moonshot's Mooncake serving system keeps cache hit rates above 90% on coding workloads, which means the real effective input cost for developers doing heavy code work is close to the $0.30 cached rate rather than the $3.00 standard rate — roughly a 4x cost reduction in practice.

A launch promotion running from July 15 through August 11 offers 10 to 30 percent bonus credits on API recharges.

Where to Access Kimi K3

Kimi K3 is available right now through four channels. You do not need to wait for the open weights release on July 27 to use it.

Kimi.com — The main web interface. Available to all registered users.

Kimi App (iOS and Android) — Mobile access with a ¥199 per month subscription tier as the entry point for K3 access.

Kimi Work — The workplace tier, aimed at professional and enterprise users.

Kimi Code — Moonshot's coding agent terminal. K3 is the flagship model in Kimi Code from day one, replacing K2.7 Code as the primary model for coding tasks. K2.7 Code remains available as a mature alternative with a 256k context window.

Kimi API — Full API access with the pricing listed above. Available to developers immediately.

Benchmarks — What the Numbers Actually Say

Moonshot published its own benchmark results at launch, and independent testing services have also started evaluating K3. Here is a clear picture of where it stands.

On Moonshot's own benchmarks:

K3 leads all tested models on Program Bench, SWE Marathon, BrowseComp, SpreadsheetBench 2, and Automation Bench. These are primarily coding, agent, and tool-use benchmarks — which reflects Moonshot's design priorities for K3.

On FrontierSWE and GDPval-AA Elo, K3 trails Claude Fable 5. On DeepSWE, it trails GPT-5.6 Sol.

On independent testing:

On Artificial Analysis and similar independent evaluation platforms, K3 lands fourth among all frontier models — behind Claude Fable 5 and GPT-5.6 Sol, and ahead of Claude Opus 4.8.

Rank	Model
1	Claude Fable 5
2	GPT-5.6 Sol
3	(tied)
4	Kimi K3
5	Claude Opus 4.8

What Moonshot admits:

In its own technical blog — which is unusual candor for a model launch — Moonshot acknowledged a noticeable user-experience gap versus Fable 5 and GPT-5.6 Sol, sensitivity issues around how thinking history is handled, and a tendency toward excessive proactiveness in agent tasks. These are real limitations worth knowing before deploying K3 in production.

Kimi K3 vs Claude Fable 5 — Direct Comparison

Capability	Kimi K3	Claude Fable 5
Parameters	2.8T	Undisclosed
Context window	1M tokens	1M tokens
Open weights	Yes (July 27)	No
Vision	Native	Native
Reasoning	Always-on	Classifier-gated
Coding (SWE Marathon)	Leads	Behind K3
Overall ranking	4th	1st
Output pricing	$15/1M tokens	$50/1M tokens
Open source	Yes	No

The most significant difference is price and openness. Claude Fable 5 costs $50 per million output tokens and is not available to download or self-host. Kimi K3 costs $15 per million output tokens and will be fully downloadable by July 27. For enterprises that need frontier-level performance but want to control their own infrastructure, this gap is meaningful.

The trade-off is overall quality and reliability. On independent benchmarks, Fable 5 still leads K3 on the aggregate leaderboard — though K3 has closed the gap significantly compared to what K2 achieved.

Kimi K3 vs Claude Sonnet 5 — Which Should Developers Use?

Claude Sonnet 5 targets the middle of the market — near-Opus performance at a lower price than Fable 5. How does K3 fit into that picture?

For most API-based development work, Sonnet 5 and K3 are competing in a similar space. K3's coding benchmarks are strong — stronger than Sonnet 5 on most tested metrics. K3's context window is significantly larger. But Sonnet 5 benefits from Anthropic's mature safety infrastructure, better documented behavior in edge cases, and strong enterprise support.

The choice depends on use case. If coding performance and context length are the priority and you are comfortable with a newer model from a less established Western enterprise vendor, K3 is worth serious evaluation. If you need predictable behavior, extensive documentation, and enterprise-grade SLAs, Sonnet 5 is the safer choice.

Why the Open Weights Matter

The open weights release on July 27 is not just a technical detail — it is what makes K3 significant beyond its benchmark numbers.

At 2.8 trillion parameters, K3 will be the largest open-weight model ever made publicly available. That means any developer, researcher, or organization with the right compute can download the entire model, run it on their own infrastructure, inspect its weights, fine-tune it on their own data, and modify it under the Modified MIT license.

This matters for the AI industry in the same way that Fireworks AI's $1 billion revenue matters — it is evidence that open-weight models at frontier scale are becoming practical alternatives to proprietary APIs. The infrastructure to run them exists. The models themselves are catching up to closed competitors. And as Linus Torvalds recently argued, the open source community's relationship with AI tools is shifting from skepticism to acceptance.

What This Means for the US-China AI Race

The timing of K3's release was not accidental. Moonshot launched K3 on July 16, 2026 — the day before the World Artificial Intelligence Conference in Shanghai, where Chinese President Xi Jinping was expected to outline Beijing's AI priorities.

K3 is not just a model. It is a signal.

China's AI development was widely considered to be running one to two years behind the US frontier as recently as 2024. The gap has closed. K3 lands at roughly the same tier as Claude Opus 4.8 — one of the strongest models that existed just a year ago — and it does so at a fraction of the cost with full open weights.

Moonshot's domestic rival DeepSeek is also expected to release an updated model in the near term, which could mean back-to-back Chinese frontier launches in the coming weeks.

For context, the Chinese AI open-source push did not begin with K3. GLM 4.5 from Zhipu AI was one of the earlier signals of this trend. K3 is a significantly larger step.

Moonshot AI's Valuation — The Business Context

K3 is also arriving at an inflection point for Moonshot AI as a company.

Moonshot closed a $500 million Series C in January 2026 at a $4.3 billion valuation — explicitly earmarked for K3 development and compute expansion. By May 2026, following additional funding, the company's valuation had reached $20 billion. Reports now suggest Moonshot is raising between $1 billion and $2 billion in a new round that would value the company at up to $31.5 billion — roughly a sevenfold increase in valuation since December 2025.

That trajectory reflects investor confidence that K3's open-weight strategy, combined with Kimi's consumer products and enterprise API, positions Moonshot as a durable frontier player rather than a one-hit wonder.

Latest Updates — Kimi K3

July 21, 2026 — Kimi Work Desktop Agent Launch

Moonshot ne Kimi Work launch ki — ek desktop agent jo K3 ko directly tumhare computer par laata hai. WebBridge feature real Chrome aur Edge browser control karta hai CDP ke through. K3 Agent Swarm directly PowerPoint aur Excel mein kaam karta hai — bina copy-paste ke. Cron scheduling se tasks schedule kar sakte ho. US, Hong Kong, aur India ka native finance data directly access hota hai. Free download available hai — paid tiers sirf swarm volume gate karte hain. Abhi sirf Apple Silicon Mac aur Windows par available hai.

July 21, 2026 — Microsoft Kimi K3 Evaluate Kar Raha Hai

Reports aa rahe hain ki Microsoft Kimi K3 ko Copilot workloads aur Azure availability ke liye evaluate kar raha hai. Open weights July 27 ko available hone ke baad Microsoft-hosted inference possible ho sakta hai. Yeh us trend ko confirm karta hai jo humne GitHub Copilot ke pehle open-weight model mein dekha tha — Chinese AI models ab mainstream developer platforms mein actively consider ho rahe hain.

July 20, 2026 — Subscriptions Pause, 7.2 Million Views

Moonshot ne new K3 subscriptions temporarily pause kar diye — launch ke sirf 48 ghante mein 7.2 million se zyada views aaye aur GPU capacity limit ho gayi. Yeh wahi pattern hai jo humne Kimi K3 chip stocks crash article mein explain kiya tha — sasta model = zyada users = zyada chips chahiye.

July 27, 2026 — Open Weights Coming

Kimi K3 ke full open weights July 27 ko publicly download ke liye available honge. Weights 1.4 TB hain 4-bit format mein bhi — locally run karne ke liye 64+ accelerators chahiye. Koi bhi consumer machine abhi yeh run nahi kar sakti. Moonshot ne vLLM ke liye KDA prefix-caching implementation bhi contribute ki hai jo weights ke saath release hogi.

Frequently Asked Questions