Kimi K3, the Largest Open-Source AI Model Ever, Is Now Free to Download

The biggest open-weight AI model in history is now a free download. Just clear 1.4 terabytes of disk space first.

Moonshot AI shipped the full Kimi K3 weights this weekend, right on schedule. The Alibaba-backed Beijing lab kept its July 27 promise.

Kimi K3, a 2.8 trillion parameter colossus

Under the hood sits a Mixture-of-Experts design with 896 experts, 16 active per token. Roughly 104 billion parameters fire on each request, out of 2.8 trillion total.

Add a 1-million-token context window, native vision, and a modified MIT license. No open model had ever reached the 3-trillion-parameter class before this one.

Benchmarks that sting OpenAI and Anthropic

Kimi K3 tops the Frontend Code Arena leaderboard, ahead of Claude Fable 5. The model also beats Claude Opus 4.8 and GPT-5.5 on coding and agentic tasks in Moonshot's own evaluations.

Only GPT-5.6 Sol and Claude Fable 5 still rank above it overall. We flagged this model's potential when its API opened in mid-July.

The open-weight escalation

Anyone with a serious GPU fleet can now compete with the American labs using a near-frontier model. OpenAI and Anthropic charge premium rates; Kimi K3 just made that conversation awkward.

The counterpunch came fast: Alibaba announced Qwen 3.8, a 2.4-trillion-parameter model with open weights on the way. Per Forbes, Chinese models already account for 60% of US token usage on OpenRouter.

The catch with Kimi K3

Self-hosting demands a multi-node cluster and 1.4 TB of combined GPU memory. The API stays the realistic entry point: $3 per million input tokens, $15 for output.

One real blemish: the hallucination rate climbed to roughly 51%, up from 39% on K2.6, per Artificial Analysis. The model answers more questions, yet knows less often when it is wrong.

Can you run Kimi K3 on your own hardware?

Not on a regular PC: the weights weigh about 1.4 TB and need dozens of GPUs across multiple nodes. Most people will use the Kimi app, the API, or cloud hosts instead.

Does Kimi K3 beat GPT-5.6 Sol and Claude Fable 5?

Not overall, where it ranks third. The model does lead on frontend code and web research, with 91.2 on BrowseComp, at roughly half the per-task cost of Opus 4.8.