What Are Open Weights? A Simple Explanation (And Why They're Eating the AI Industry)

What Are Open Weights? A Simple Explanation (And Why They're Eating the AI Industry)

If you've seen "open weights" attached to models like DeepSeek, GLM, Kimi, or Llama and wondered how that differs from "open source" — or why a company would give away its most valuable asset — this explainer is for you. No jargon, no math degree required.

The one-sentence version

Open weights means the trained "brain" of the AI model — the billions of numbers that encode what it learned — is downloadable, so anyone can run, modify, or host it themselves. The recipe (training code and data) usually stays private; only the finished product is shared.

A kitchen analogy

Think of a model as a finished dish:

• Closed model (GPT-6, Claude): A restaurant meal. You can eat it, but only at their restaurant, at their price, and the chef decides the recipe forever. If they close, change the recipe, or raise prices — you have no alternative.
• Open weights (DeepSeek V4, GLM 5.3, Llama): The finished dish, packaged, with permission to reheat it, serve it in your own restaurant, or tweak the seasoning. You don't get the original recipe, but you have the dish itself — permanently.
• True open source: The full cookbook, including sourcing and training steps. This is rare — a few models ship training code and data, most don't.

What "weights" actually are

Inside every large language model are billions of adjustable numbers called parameters or weights. During training, the model adjusts these numbers billions of times so its predictions get better. After training, the weights are the model — everything it "knows" is encoded in them.

When a lab "releases open weights," it publishes a file (often tens to hundreds of gigabytes) containing those numbers. Download it, and you can run the model on your own hardware: a rented GPU, a server rack, or in some cases a laptop.

Why would a company give this away?

It sounds like giving away the store, but the economics work surprisingly well:

1. Ecosystem pull. Developers build on open models first because there's no lock-in. That familiarity drives adoption of the lab's paid, hosted API.
2. Distribution without a sales team. Every cloud provider, marketplace, and startup hosting your model is marketing it for free.
3. Talent and trust. Researchers want to work where their work is public. Enterprises trust models they can audit.
4. It doesn't cost inference revenue they'd otherwise get. Many buyers can't self-host anyway — they'll use the hosted API regardless, and the open weights function as a quality guarantee: "you can leave anytime, so we have to stay good."

DeepSeek's recent $1B in annual revenue — earned largely on a model anyone can download — is proof the model works.

What open weights does NOT give you

This is where people get tripped up:

• Not the training data. You get the finished model, not what it learned from.
• Not the training code, usually. Reproducing it from scratch remains hard and expensive.
• Not always "free for any use." Licenses vary wildly: some are MIT/Apache (do anything), others restrict commercial use above certain revenue thresholds, or require naming the origin model. GLM-5.3, for instance, ships a custom license that mandates a security review only for companies making over $10B/year running model-as-a-service.
• Not free to run. Weights are free; the GPUs to serve them are not. Self-hosting only beats renting API access at significant scale.

Why open weights are winning right now

Three forces converged in 2026:

1. Quality parity. Open-weights models like GLM-5.3, Kimi K3, and DeepSeek V4 now match or beat closed flagships on coding and agentic benchmarks. The "open models are a year behind" rule of thumb is dead.
2. Price collapse. When weights are public, anyone with cheap GPUs can serve them — and that competition drives prices toward actual cost. Frontier-class work that cost $30/M tokens at retail serves for under $1 when bought on open markets.
3. No lock-in. After a year of price hikes, deprecations, and forced migrations in the closed world, developers are weary. An open-weights model can't be deprecated out from under you.

How to actually use open-weights models

You have three options, from easiest to most control:

1. Call them via API through a marketplace. Zero ops; you get open-weight economics without managing servers. On Qubax, open models like DeepSeek V4 Pro and GLM 5.3 are available at open-market rates that track wholesale cost — often 60–80% below standard retail.
2. Self-host on rented GPUs. Best at scale, or when you need privacy guarantees. Tools like vLLM and SGLang serve open models efficiently.
3. Run locally. Small open models (2–30B parameters) now run on laptops and phones — great for privacy-sensitive or offline work.

The takeaway

Open weights is the middle ground that turned out to be the winning ground: not full open source, but enough freedom that no single company controls the model. Once weights are public, the model's price converges toward the cost of running it — and the company that serves it cheapest wins. That's the race the AI industry is actually running now.

Want to compare open-weight models against closed flagships at live, open-market prices? Browse the full model list on Qubax AI — DeepSeek, GLM, Kimi, Llama, GPT, and Claude side by side, at prices that track real cost.

FAQ

Is open weights the same as open source?

No. Open weights publishes the trained model parameters; open source also includes training code and data. Most "open" models today are open weights only, under licenses that range from fully permissive (MIT, Apache 2.0) to custom terms with restrictions.

Can I use open-weights models commercially?

Usually yes, but check the specific license. Some (Llama variants historically, GLM-5.3) have conditions like revenue thresholds or naming requirements. MIT and Apache 2.0 licensed models have essentially no restrictions.

Are open-weights models as good as GPT or Claude?

For most tasks, yes — the best open models (DeepSeek V4, GLM-5.3, Kimi K3) now match closed flagships on coding, reasoning, and agentic benchmarks. Closed models still lead at the very frontier on some specialized tasks.

Do open weights cost anything to run?

The weights themselves are free, but you need compute to run them. That's why most developers use hosted APIs — marketplaces like Qubax offer open-weights models at open-market prices without the server management.