The Biggest Free AI Model Yet Arrives July 27. Here's What 'Open Weights' Actually Means

Moonshot AI has put a date on something the industry has been waiting for: on July 27, the company plans to publish the weights of Kimi K3, the model that topped a major coding leaderboard days after its launch. Anyone will be able to download it. That sounds like a bigger deal than it is for most people, and also like a smaller one than it is for the industry, so it is worth unpacking both.

Start with the term. A model’s weights are the billions of numbers it learned during training, effectively everything it knows, stored as a very large file. When a lab releases them, you can run the model on your own hardware, inspect it, adapt it, and use it without asking permission or paying per question. When a lab keeps them, as OpenAI and Anthropic do with their flagship models, you can only reach the model through their service. That is the entire difference between “open weights” and “closed”.

Now the practical caveat, and it is a large one. K3’s weights are reported to be roughly 594 GB in their full form. Community versions that compress the numbers, a process called quantisation, will likely bring that down to somewhere around 300 to 400 GB. Either way, this is not a model you run on a laptop. It needs serious server hardware, so in practice the immediate beneficiaries are companies, research groups, and hosting providers rather than individuals. What individuals get is second-hand: cheaper access, because anyone can now host the model and compete on price.

What is behind it: Moonshot is playing the same game DeepSeek played, and it works. Giving the weights away costs a Chinese lab relatively little upfront and buys enormous distribution, credibility, and pressure on Western labs that charge per use. It also makes the model very hard to ban or cut off, since copies spread the moment it is published. Note the honest asterisk though. K3 leads on coding and agent tasks but ranks around ninth on general conversation, so it is a specialist, not an all-round replacement for the best paid models.

One related date is being widely misreported, so here is the accurate version: July 24 is not a DeepSeek V4 launch. It is the day DeepSeek retires its older model names, after which requests to those names stop working. Useful if you or your developer built anything against them, irrelevant otherwise.

What this means for you: if you use AI through a chat app, nothing changes on July 27, and you do not need to do anything. What you may notice over the following months is prices drifting downward, because every strong free model makes it harder for paid ones to justify their rates. If you run a small company and are paying meaningful money for high-volume, repetitive AI work, this is the moment to ask a technical colleague whether an open model hosted elsewhere would do the same job for less. The honest answer is often task-dependent, and worth measuring rather than guessing.

Sources

• Crypto Briefing: Kimi K3 launches with 2.8 trillion parameters, open weights dropping July 27
• VentureBeat: Moonshot AI releases Kimi K3, the largest open-source model ever
• DeepSeek API docs: model retirement notice