Moonshot AI's Kimi K3 open weights went public on July 26, 2026, a day ahead of the July 27 target the company had promised. The headline numbers are real: a 2.8 trillion parameter mixture of experts model, 104 billion active parameters per token, native vision, and a 1 million token context window. What most of the coverage is getting wrong, or simply not checking, is the license attached to those weights. It is not the "Modified MIT" license that shipped with Kimi K2. It is a new document with a specific commercial threshold, and that difference matters more than another benchmark chart.

Kimi K3 is a 2.8T parameter MoE model built on Kimi Delta Attention and Attention Residuals, activating 16 of 896 experts per token (104B active). It ships with a 1M token context window, native multimodal input, and a technical report claiming roughly 2.5x better scaling efficiency than Kimi K2. Moonshot's own report places K3 behind Claude Fable 5 and GPT-5.6 Sol on overall intelligence, but ahead of every other open or closed model tested, including Claude Opus 4.8 and GPT-5.5, across most coding and agent benchmarks. The weights are downloadable now from Moonshot's Hugging Face organization, but the license is a custom "kimi-k3" document, not a plain MIT variant, and it gates commercial inference use above $20 million a year in revenue.

The rollout had two stages. Kimi K3 first went live as a hosted product on July 16, 2026, through the Kimi app, Kimi Code, and the API. Moonshot said at the time that full weights would follow by July 27. They landed a day early, at roughly 7:30 PM EDT on July 26, alongside the technical report.

The core specs, drawn from Moonshot's own technical report:

The jump from 384 to 896 experts, with the active count only doubling from 8 to 16, is the part worth sitting with. Moonshot is not just adding parameters, it is making the model more selective about which of those parameters fire on a given token, which is where the 2.5x scaling efficiency claim comes from rather than from raw scale alone.

Every article covering K3 pulls two or three numbers from Moonshot's launch blog. The technical report itself contains a fuller picture across coding, general agent work, and vision tasks, tested at maxed-out thinking effort. Pulled directly from the report and cross-checked against the independently tracked Arena results:

Coding

General and visual agents

K3 leads Fable 5 and GPT-5.6 Sol on several individual rows, including BrowseComp and FrontierSWE, but Moonshot's own summary is blunt about the overall gap: K3's general user experience still trails both of those models. That kind of self-graded honesty is unusual enough that it is worth taking at face value rather than assuming a hidden angle.

One benchmark detail is easy to miss and changes how you should read the context window number. K3's reported BrowseComp score of 91.2 uses context compaction at 300K tokens, following the same strategy Anthropic uses on Claude model cards. Run with the full 1M window and no compaction, the same benchmark scores 90.4. A bigger context window does not automatically beat a smaller, better-managed one, and that gap is worth remembering before you architect an agent around "just dump everything into the million tokens."

Ask five sources what license Kimi K3 ships under and four will say "Modified MIT," because that is what Kimi K2, K2.6, and K2.7 Code used. It is a reasonable prior. It is also wrong for K3.

The actual license on the Hugging Face repository is listed as a custom document, not a standard MIT variant. According to the license text itself and independent confirmation from open model researcher Nathan Lambert, who read the file directly, Kimi K3's terms split into two separate commercial gates:

Compare that to K2's actual license text, which only required the attribution display above 100 million MAU or $20 million monthly revenue, with no separate MaaS revenue gate at all. K3 adds a distinct, lower commercial threshold aimed specifically at anyone who wants to resell K3 as an API. If you are a startup building a product with K3 inside it, you likely clear the bar comfortably. If your plan is to compete with Moonshot's own API business, the $20 million a year MaaS threshold is the number your legal team needs, not "Modified MIT," and that is the detail every other roundup published this week either got wrong or hedged into vagueness. Read the LICENSE file in the repository yourself before you ship anything commercial on top of K3, because the terms genuinely changed generation to generation.

The claim is repeated everywhere: largest open weight model in history. It is directionally true and worth a caveat. Among 2026's tracked open frontier releases, K3's 2.8 trillion parameters beats DeepSeek V4 Pro at 1.6 trillion and MiMo V2.5 Pro at roughly 1 trillion by a wide margin. That is a real, checkable comparison. What it is not is an independently audited record against every open weight model ever released; it is the largest among the models currently being tracked and benchmarked in the same news cycle, which is a narrower and more honest claim than the headline version.

A second claim making the rounds, that Moonshot went from launch to open weights in "under 24 hours," does not survive a timeline check either. K3's API launched July 16. The weights landed July 26, roughly ten days later, one day ahead of Moonshot's own stated July 27 target. Beating your own deadline by a day is a genuinely fast turnaround for a 2.8T model with a technical report attached. It is not the same claim as a 24 hour launch to open source, and the two get conflated in social posts often enough that it is worth separating them before you repeat either one.

The API is priced at $3 per million input tokens on a cache miss and $15 per million output tokens, with a much cheaper $0.30 per million tokens on a cache hit. Moonshot's disaggregated Mooncake inference architecture reportedly holds cache hit rates above 90% on typical coding workloads, which means most real coding traffic lands closer to the $0.30 figure than the $3 headline rate. For budgeting purposes, the sticker price is not the number that matters if your workload is repetitive tool calling and file reading, which most agentic coding sessions are.

Running it yourself is a different budget entirely. Moonshot recommends a minimum of 64 accelerators for serving, and the safe planning assumption for self-hosting is an 8x H100 80GB cluster or larger, even with the MXFP4 quantization applied at training time. This is not a model you point Ollama at on a desktop GPU. If your actual need is local, single-GPU inference, K2.7 Code remains the realistic Moonshot option, not K3.

Two things are happening around this release that a pure spec sheet leaves out.

First, the market reaction was immediate and large. Z.ai's Hong Kong-listed stock dropped as much as 30% and MiniMax fell as much as 16% on the news of K3's initial launch, with Alibaba down 4%, according to Bloomberg's reporting picked up by Quartz. Moonshot's own daily revenue reportedly grew by a factor of at least six since K3 launched, with annual recurring revenue reaching $300 million in June, up from $200 million in April, and the company is said to be raising at a $50 billion valuation ahead of a possible Hong Kong IPO.

Second, the release landed inside an active US policy fight. White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot of training K3 on restricted Nvidia chips and of running large-scale distillation against US models, including Anthropic's Fable line, an allegation Moonshot has not publicly responded to. That sits alongside an earlier, separate accusation from Anthropic in February that Moonshot used millions of Claude conversation logs to train prior Kimi models. Neither claim has been independently verified as of this writing, and Moonshot has not confirmed or denied either one on the record. On the other side of the debate, AMD, Vercel, and Ollama publicly signed a letter arguing for keeping open weight models, including Chinese-origin ones, broadly available, while Anthropic and OpenAI have reportedly pushed for tighter restrictions. This is a live, contested policy question rather than a settled one, and it is worth reading K3 coverage from any single outlet with that in mind.

Is Kimi K3 actually open source, or just open weights? Open weights. Moonshot is releasing the trained parameters, not the training code, training data, or a full reproduction pipeline. For most builders that is enough to self-host and fine-tune, but it is a narrower claim than "open source."

What license does Kimi K3 use? A custom document Hugging Face lists as "kimi-k3," not a standard MIT license. It permits most research, internal, and smaller commercial use, but requires a separate commercial agreement with Moonshot once your Model-as-a-Service use passes $20 million a year in revenue, and requires attribution display once a product built on it passes 100 million monthly active users or $20 million a month in revenue.

Can I run Kimi K3 on a single GPU? No. Moonshot recommends at least 64 accelerators for serving, and realistic self-hosting starts around an 8x H100 80GB cluster even with MXFP4 quantization. For single-GPU local inference, Kimi K2.7 Code is the practical option.

Does Kimi K3 beat Claude Fable 5 and GPT-5.6 Sol? Not overall, by Moonshot's own account. K3 trails both on general intelligence and user experience but leads or matches them on several individual coding and agent benchmarks, including BrowseComp and FrontierSWE.

Is Kimi K3 really the largest open weight model ever released? It is the largest among 2026's actively tracked open frontier models, beating DeepSeek V4 Pro and MiMo V2.5 Pro by a wide margin. That is not the same as an independently audited record against every open weight model in history, which is a claim nobody has actually verified.