The models reshaping the AI market this month are open weight. Understanding what that means, how many doors lead to them, and what they do to a business’s cost structure explains why the shift matters.

The weights are the model

A large language model is, at its core, a very large file of numbers. Those numbers are the weights: the learned parameters that encode everything the model knows how to do. Training produces them at enormous cost. Once they exist, running the model is comparatively cheap.

“Open weights” means the lab publishes that file. Anyone can download it, run it on their own hardware, fine-tune it for their own purposes, and build products on top of it. The lab gives up the ability to meter access or revoke it.

This differs from open source in the strict sense. Open source software includes the recipe: the training code, the data, the full pipeline. Open-weight labs publish the finished weights and a technical report, and keep the training data private. For practical purposes, the weights are what matter. You can use the model fully without the recipe.

It differs completely from closed models. GPT, Claude, and Gemini are accessed through an API or a chat window. The weights stay on the company’s servers. You rent capability by the token, and the company can change pricing, deprecate versions, or cut you off. With open weights, the file is on your disk. It owes its maker nothing further, and neither do you.

Kimi K3, the live example

Moonshot AI announced Kimi K3 on July 16. The model went live immediately through the Kimi app, Kimi Work, Kimi Code, and the Kimi API. The full weights, all 2.8 trillion parameters of them, arrive as a public download by July 27. It will be the largest open-weight model ever released, and the release comes with the technical report describing architecture, training, and evaluations.

The two-stage launch is a great way to understand how models work because it shows what “open” means in practice. For the first eleven days, K3 works like any hosted model: you use it through Moonshot’s products and pay by the token. On the 27th it becomes a folder on anyone’s disk. Every question about pricing, access, capacity limits, and jurisdiction changes on that day.

Where to get open-weight models

Hugging Face is the primary distribution hub. Nearly every open-weight release lands at huggingface.co within hours, weights and documentation and community discussion together. Labs publish under their organization pages; Moonshot’s is huggingface.co/moonshotai, alongside DeepSeek, Qwen, GLM, Llama, and Mistral.

GitHub carries the lab repositories: license text, inference code, deployment guides. Checking a lab’s GitHub organization is the fastest way to confirm a release is real rather than rumored.

ModelScope, Alibaba’s hub, is the main distribution point inside China and mirrors most major Chinese releases. Useful to know when Hugging Face access is slow or when a release appears there first.

Ollama and LM Studio are free desktop apps that make local running trivial. Install, pick a model from the catalog, and it downloads and runs on your machine. No account, no API key, no data leaving your computer.

Inference providers host open-weight models and sell API access: OpenRouter aggregates dozens of hosts behind one API, while Together, Fireworks, Groq, and DeepInfra compete directly on price and speed. Because they all serve the same free file, margins compress toward zero and the savings pass through.

Cloud platforms carry open-weight models inside enterprise procurement channels. AWS Bedrock, Azure AI Foundry, and Google Vertex all offer major open models with the compliance paperwork, billing integration, and security posture large organizations require.

The labs themselves run free chat interfaces: kimi.com, chat.deepseek.com, chat.qwen.ai. No waitlist, no invoice.

Hardware, latency, and the speed question

The full K3 is datacenter-scale. Practical deployment starts around 1.5 terabytes of memory, and Moonshot recommends supernode configurations of 64 or more accelerators. Almost nobody self-hosts a model that size, and almost nobody needs to. Within days of any major release, the community produces quantized versions, compressed variants that trade some capability for the ability to run on serious workstations, and the smaller siblings in each model family run acceptably on a recent laptop.

Latency is where self-hosting can be the better option. A model on your own hardware answers without a network round-trip, without rate limits, without queues, and without capacity walls. That last one is current: Moonshot has already stopped taking new API customers because demand outran its serving capacity. A hosted model can be full. A downloaded one never is.

The tradeoff runs the other way on raw speed for big models. A giant model squeezed onto modest hardware produces slow tokens, and specialized inference providers running custom silicon will beat your server rack on throughput. The practical answer is to right-size: match the smallest model that does the job to the hardware you have, and rent the big ones by the token when the task demands it.

What this means for small businesses

The immediate implication is price. Hosted open-weight models cost a fraction of the closed incumbents because a dozen providers compete to serve the same file. K3’s own API launched at 30 cents per million cached input tokens. For a small firm, model spend drops from a budget line to a rounding error.

The deeper implication is ownership of the routine layer. A capable mid-size model on a single decent machine handles drafting, summarization, classification, translation, and first-line customer support. That covers most of what a small business actually asks of AI, with no per-seat licensing, no usage meter, and no monthly bill that scales with success.

For firms that handle confidential material, law, accounting, health, finance, local running means client data never leaves the premises. That single property settles the buy-versus-build question for a lot of professional services.

What this means for enterprises

Data residency and compliance. Self-hosted weights run in whatever jurisdiction the enterprise chooses. Regulated industries that could never send data to a foreign API can run the same intelligence inside their own perimeter.

The end of vendor lock-in. Weights on your infrastructure survive the vendor’s pricing changes, its deprecation schedule, its outages, and any diplomatic rupture between the country that made the model and the country that runs it. A downloaded model is a capital asset. An API subscription is a dependency.

Fine-tuning on proprietary data. Open weights can be trained further on internal documents, code, and processes without any of that material leaving the building. The result is a model shaped to the business that the business fully controls.

Predictable cost at scale. Heavy API usage compounds; self-hosted inference is an infrastructure cost that behaves like one. Past a certain volume, owning the serving stack is simply cheaper, and the crossover point keeps dropping as the models get more efficient.

The caveats are real and manageable. Open weight is a narrower promise than OSI open source, which matters to compliance teams with strict definitions. License terms vary by release and the exact text appears only when the weights do; Moonshot’s prior models shipped under a modified MIT license permitting commercial use, and K3 is expected to follow, but expected is doing work in that sentence until the 27th. And enterprises weighing Chinese-origin models will face procurement and security review questions that have more to do with politics than with the file itself. The file, once downloaded, behaves identically everywhere.

The hybrid pattern is the sensible default. Evaluate through a hosted API, where switching costs are near zero. Self-host what proves out, where the compliance and cost advantages live. The same weights serve both stages.

A twenty-minute start

Install Ollama. Run a small Qwen or DeepSeek variant from its catalog. Ask it the same questions you ask your paid chatbot. The point of the exercise is less the answers than the realization of what just happened: a frontier lab’s work product, running on your laptop, for free, with no one’s permission required.