Kimi K3: Everything You Need to Know, and How to Actually Access It

The Verdict

Kimi K3 is a real frontier model, not a press release. The weights are live, the benchmarks are credible where independently verified, and the architecture is genuinely novel. The distillation allegations are serious, partially documented, and — on the K3-specific timeline — not yet proven.

You need to hold both of those facts at once.

What It Is

Moonshot AI has launched Kimi K3, a 2.8-trillion-parameter mixture-of-experts model billed as the world's first open "3T-class" AI system. With a 1-million-token context window, native vision capabilities, and a new architecture built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), K3 is positioned as a direct challenger to Claude Fable 5 and GPT-5.6 Sol at roughly half the price.

The 2.8 trillion figure is real but requires translation. It is a Mixture-of-Experts design with 2.8 trillion total parameters and roughly 104 billion activated per token. Translation: you are not running 2.8 trillion parameters per forward pass. You are running roughly 104 billion. The headline describes the weight file; the active-parameter count describes your inference bill.

K3 uses 896 experts with just 16 active per token — meaning roughly 50 billion parameters per forward pass. This extreme sparsity is what makes a 2.8T model practically servable. (Moonshot's official figure is 104 billion active; independent estimates cluster around 50 billion. Moonshot has not published a single official active-parameter definition alongside the weight release.)

KDA is a hybrid linear attention mechanism interleaving linear-attention layers with periodic full-attention layers in a 3:1 ratio. The model accepts text, image, and video input, always reasons, and now exposes low, high, and max reasoning effort rather than max only.

The Benchmarks Worth Believing

K3 scores within 3 points of Claude Fable 5 on the Artificial Analysis Intelligence Index (57 vs. 60) and beats it outright on the Frontend Code Arena leaderboard with 1,679 Elo — the top open-weight result, with GLM-5.2 next at 51.

Moonshot reports strong coding results: 67.5 on DeepSWE (corroborated on the official leaderboard), 77.8 raw pass rate on ProgramBench, 88.3 on Terminal-Bench 2.1, 81.2 dominance on FrontierSWE, and 42.0 on SWE Marathon.

The cyber evaluations complicate the story. The UK Artificial Intelligence Security Institute and U.S. Center for AI Standards and Innovation conducted a joint evaluation and found K3 performing significantly below top frontier models on preliminary cyber evaluations — ExploitBench came in at 32% versus roughly 76% for average top U.S. models. Before you read too much into that gap: the U.S. models were tested with safety measures disabled, while K3 was only evaluated under limited conditions due to hosting constraints. The number is real. The framing is contested.

The cyber gap matters to the distillation debate. A model trained primarily on curated distillation data tends to be strong on distilled tasks and weak on harder, less queryable capabilities.

The pattern fits. It does not constitute proof.

The Distillation Controversy

Two separate things are being conflated here. Keep them separate.

What is documented. Anthropic identified industrial-scale campaigns by DeepSeek, Moonshot, and MiniMax to illicitly extract Claude's capabilities through over 16 million exchanges via approximately 24,000 fraudulent accounts. Moonshot AI's portion: over 3.4 million exchanges targeting agentic reasoning and tool use, coding and data analysis, computer-use agent development, and computer vision. That disclosure came in February 2026. It is documented. It is not disputed by Moonshot in any technical detail.

What is alleged but unproven. White House OSTP Director Michael Kratsios claimed that Moonshot distilled Anthropic's Fable model to develop K3, using "a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection." Kratsios provided no details on how the U.S. government learned this. An OSTP director posting on X is not a legal filing. It is not technical proof.

The timeline is the evidentiary problem. Only about 16 days separate Fable 5's July 1 return to general availability from K3's July 17 launch. You do not train a 2.8-trillion-parameter model in two weeks on data from a newly available API. The earlier distillation campaign is real. Its direct connection to K3 specifically has not been established with evidence.

Treasury Secretary Scott Bessent warned within hours that sanctions and Entity List designations "will be on the table." As of late July 2026, no enforcement action has actually been imposed — every consequence in circulation is floated, threatened, or under investigation.

The separate export-control allegation is more legally actionable. Kratsios said Moonshot acquired restricted Blackwell-generation servers in Thailand. Hardware sanctions enforcement operates on physical evidence. That thread is worth watching more closely than the distillation narrative.

How to Access It

The weights are live and ungated. Open weights shipped on July 27, 2026 — 96 shards and about 1.56 TB on Hugging Face, under a custom Kimi K3 License rather than MIT. Read that license file before commercial deployment. The license permits use, modification, distribution, and commercial sale — but a model-as-a-service business earning over $20 million in any 12 months needs a separate agreement, and products with over 100 million monthly users must credit Kimi K3 on screen.

Your realistic access options:

• Consumer/product layer: kimi.com, Kimi Code, Kimi Work. Data routes through Moonshot's servers under Chinese law. Budget and deadline teams who do not touch sensitive data: this works today.
• Hosted API: Kimi K3 is available through Moonshot's API as kimi-k3 and through OpenRouter as moonshotai/kimi-k3. It costs $3.00 input / $0.30 cache hit / $15.00 output per 1M tokens, has a 1M-token context window, and accepts text, image, and video input. Same data-sovereignty caveat applies.
• Third-party hosted inference: Together AI and Modal both shipped day-0 hosted access on the weights. For teams needing non-Moonshot infrastructure without spinning up their own cluster, this is the practical path.
• Self-hosted: Moonshot recommends supernode configurations with at least 64 accelerators for deployment. At 1.56 TB of quantized weights, this is a datacenter job.

One cost reality check before you budget: K3 uses maximum reasoning effort by default, with additional effort settings to be released later. Always-on maximum reasoning means high output-token consumption. The $15/million output rate can close the gap with closed competitors faster than the headline input price suggests. Run your own token math on real workloads before committing.

What This Actually Means

Moonshot is reportedly raising between $1 billion and $2 billion in a new round that would value it at up to $31.5 billion — roughly 50% higher than the $20 billion valuation in May 2026. That would mark a seven-fold increase since December 2025, when Moonshot was worth just over $4 billion.

This is not purely a technical story. It is a capital story wearing a benchmark costume.

For operators: K3 is the most capable open-weight model available by a meaningful margin. The data-sovereignty question is real and non-negotiable for regulated industries — the open-weight release addresses it for teams with cluster infrastructure, and Together AI / Modal address it for everyone else. The distillation controversy matters to IP and regulatory debate. It does not change the model's performance on your actual codebase. Run it, measure it, decide with your own numbers.

The consensus view six months ago was that open-weight models were a generation behind closed frontier. That consensus is now expensive to keep believing.