Technology & AI
Open-Weight Models
The short version
Training an AI model produces its weights: billions of numbers that hold everything it learned. An open-weight model publishes those numbers, so anyone can download the model and run it on their own hardware. A closed model (Claude, Codex, Grok, ChatGPT, Gemini) keeps its weights private. You only get to use it through the company's app or API.
Open weight means you get the finished model to keep. You usually do not get the recipe.
Open weight vs closed vs fully open source
| Closed | Open weight | Fully open source | |
|---|---|---|---|
| How you use it | API or app only | Download and run it yourself | Download and run it yourself |
| Trained weights | Private | Published | Published |
| Training code | Private | Sometimes | Published |
| Training data | Private | Usually private | Published or fully documented |
| License | Terms of service | Often has limits (commercial caps, use rules) | Permissive |
| Examples | Claude, Codex, Grok, GPT, Gemini | Llama, Mistral, DeepSeek, Qwen, Gemma, gpt-oss, Kimi | OLMo, Pythia |
Most models people call "open source" are really open weight. Without the data and training code, you can use and adapt the model but you can't rebuild it from scratch.
Why it matters
- Privacy: run it on your own machine, offline, so your data never leaves.
- Cost: no per-call API fees once you have the hardware.
- Customization: fine-tune it on your own data.
- No lock-in: nobody can change, reprice, or retire the model on you.
The tradeoffs
- You supply the hardware, and that gets expensive fast for big models.
- Guardrails are easier to strip out. Once weights are public, safety tuning can be fine-tuned away. This is the main policy worry.
- Licenses vary. "Open" doesn't always mean free for any use.
- The very best closed models are often still ahead on the hardest tasks, though the gap keeps shrinking.
Size: smaller to run, but not always small
The open models most people actually use are small (roughly 1B to 70B parameters) because they have to fit on hardware people own: a laptop, a gaming GPU, a phone. Tricks like quantization (storing weights at 4-bit instead of 16-bit) shrink them further.
The frontier is a different story. The biggest open models now rival closed ones in size by using a mixture-of-experts design, where only a slice of the model runs for each word. They are open, but you need a data center to run them.
Biggest open-weight model as of October 2026: Kimi K3 from Moonshot AI. It has 2.8 trillion total parameters (about 104 billion active per token) and a 1 million token context window, and its weights came out in July 2026. Runner-up is Qwen's 2.4 trillion parameter model. Mistral Large 4 (1 trillion) has weights scheduled for October 27, 2026.
How people run them
- Get them: Hugging Face.
- Run locally: Ollama, LM Studio, llama.cpp.
- Serve at scale: vLLM.