The Largest Open Model Ever Is Free. Can You Run It?

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

3 things that happened while you were busy

1.  The largest open-weight model in history is now downloadable.

Moonshot AI shipped the full weights of Kimi K3 on July 27, a day early: 2.8 trillion parameters across 96 shards, about 1.56 terabytes. It comfortably takes the size crown from DeepSeek's 1.6T V4-Pro. It is a Mixture-of-Experts design, so only 104 billion of those parameters fire per token, which keeps the actual compute cost far below what the headline number suggests.

2.  The independent benchmarks are real, and strong.

Unlike most launches, this one arrived with third-party data. On the Artificial Analysis Intelligence Index, K3 scored 57, the top open-weight result, trailing only closed models Claude Fable 5 at 60 and GPT-5.6 Sol at 59, and edging out Claude Opus 4.8. On the blind Frontend Code Arena, developers ranked it first overall. A model within three points of the closed frontier is now free to download.

3.  The license quietly changed, and most coverage missed it.

Everyone called it "Modified MIT," the license Kimi K2 shipped under. It is not. K3 uses a new custom Kimi K3 License with a commercial-scale threshold. It is still free for the vast majority of commercial users, but if you are building at large scale, read the actual terms before you build on it. The benchmark chart got the headlines; the license is the part that affects your business.

Free to download is not the same as free to run.

The gap.  This is the part the celebration skips. At 1.56 terabytes, K3 does not fit on a single machine or even a single GPU pod. Self-hosting it means provisioning a multi-node GPU cluster, which is out of reach for hobbyists and most small teams. "Open weights" and "runs on your laptop" are very different promises, and only one of them is true here.

The middle path.  So how do people actually use it? Through inference providers. Together AI and Modal shipped day-zero hosted access, meaning you rent K3 by the token just like a closed model, with the difference that no single company controls it. The practical setup most teams land on: hosted API for the giant model, self-hosted weights only for smaller models where the hardware math works.

The price signal.  The pressure this puts on closed models is not really about self-hosting. It is pricing. K3's hosted API runs $3 per million input tokens, the same headline rate as Claude Sonnet 5, for near-frontier quality. When a downloadable model matches a closed one on both benchmarks and price, the argument that frontier capability requires a frontier budget gets much harder to make.

The takeaway.  For almost everyone reading this, K3 is not something you host. It is something that makes the model you already pay for cheaper and better, by giving every provider a credible frontier-class alternative to compete against. That is the real gift of open weights: not that you run them, but that their existence disciplines the whole market's prices. The prompt below helps you turn that into an actual stack decision.

What people are actually watching and sharing

The pelican test.  Developer Simon Willison ran K3 through his signature "pelican riding a bicycle" SVG benchmark, the informal vibe-check the community actually trusts more than official charts. The forcing function is the point: a benchmark you can run yourself beats one you have to take on faith.

Microsoft's shopping list.  Reporting attributed to The Information says Microsoft is evaluating Kimi K3 for possible Copilot workloads and preparing Azure availability. A Chinese open model inside Microsoft's stack would have sounded unthinkable a year ago.

Capacity, again.  Demand in the first 48 hours pushed close to the limits of Moonshot's capacity, the same compute-constraint story that shadowed the K3 launch two weeks ago. The weights are free; the servers to run them are not.

A dense open-weight week.  K3's weights landed the same week as DeepSeek V4's stable release and an open-weight policy letter signed by Nvidia, Meta, and Microsoft. The timing was almost certainly not a coincidence.

Design your AI stack without wasting money

Kimi K3 proves the menu is now closed premium, hosted open, and self-hosted, and picking wrong on any of them either overspends or overcomplicates. Paste this into Claude or ChatGPT and get a setup matched to your real budget and skills, with the self-hosting hype filtered out. Most people need far less than the internet implies.

Be honest about the technical-ability line, since it changes everything: "none" points you to simple apps, "can run servers" opens up cheaper options with more setup. Advanced tip: ask it to "redo this assuming my budget is half" to see what you would actually give up, which is the fastest way to find your real priorities.