What is an open-weight AI model — and why you can't run a trillion parameters at home
An open-weight model is an AI model whose weights, the numbers learned during training, are published so anyone can download them and run the model on their own computers or servers. It is not the same as open source: training data and training code are usually not released, and the license can limit use. Meta's Llama license requires separate permission for services with more than 700 million monthly users, while Apache 2.0 and MIT licenses allow commercial use almost without restriction. Size is the other limit: a 1-trillion-parameter model needs roughly half a terabyte of memory even at 4-bit precision, so models like Mistral Large 4, released in October 2026, are in practice for data centers
The three lines
- Definition — downloadable trained weights you can run yourself; data and code are separate questions
- Licenses — Apache 2.0 and MIT are near-free; Llama needs permission above 700M monthly users
- Reality — parameters × bits = memory; 1T parameters is ~0.5 TB at 4-bit, beyond a PC
Key questions
- Open weight vs open source AI
- **Open weights give you the finished dish; open-source AI also gives you the recipe.** | Released | Open-weight | Open-source AI (OSI) | |---|---|---| | Weights | Yes | Yes | | Inference code | Usually | Yes | | Training code | Usually not | Yes | | Training data information | Usually not | Enough to rebuild | | Use restrictions | License-dependent | Free to use, modify, share |
- Can open-weight models be used commercially
- **Read the license first.** | License | Commercial use | Key terms | |---|---|---| | Apache 2.0 | Yes | Notices, patent grant | | MIT | Yes | Copyright notice | | Llama Community License | Conditional | Permission above 700M monthly users; acceptable-use policy | | Custom | Varies | Revenue, use or region limits possible |
- How much memory to run an LLM locally
- **Roughly parameters × bytes per number.** | Model | 16-bit | 8-bit | 4-bit | |---|---|---|---| | 8B | ~16 GB | ~8 GB | ~4 GB | | 70B | ~140 GB | ~70 GB | ~35 GB | | 501B | ~1 TB | ~500 GB | ~250 GB | | 1T | ~2 TB | ~1 TB | ~500 GB | Weights only; long conversations need more.
Plenty of AI models call themselves "open." What is actually open varies a lot. On October 6, 2026, Mistral previewed its 1-trillion-parameter Large 4 and promised to release the weights by the end of October. Reflection AI's Beam and China's DeepSeek, GLM and Qwen models are also described as "open-weight." Here is what that term means, how it differs from open source, and the catch behind "anyone can use it."
1. What weights are, and what gets released
An AI model is a vast set of numbers. Training adjusts them; once fixed, they are the weights, or parameters. "One trillion parameters" means a trillion such numbers.
| Component | What it is | Analogy |
|---|---|---|
| Weights | Numbers learned in training | The finished dish |
| Inference code | Software that loads weights and computes answers | Reheating instructions |
| Training code | Software that produced the weights | The recipe |
| Training data | Text and images the model learned from | Ingredients and sourcing |
An open-weight release publishes the weights (and usually inference code). You can install the model on your own server, fine-tune it on your data and run it offline. Closed models such as ChatGPT, Claude and Gemini run only on the vendor's servers; you send requests through an API.
2. How it differs from open source
In October 2024 the Open Source Initiative published its Open Source AI Definition 1.0. To qualify, a model must release weights, training code and enough information about training data to build a substantially equivalent system, with freedom to use, study, modify and share.
| Closed | Open-weight | Open-source AI (OSI) | |
|---|---|---|---|
| API use | Yes | Yes | Yes |
| Download weights | No | Yes | Yes |
| Run and fine-tune on your servers | No | Yes | Yes |
| Training code | No | Usually no | Yes |
| Training data information | No | Usually no | Yes |
| Restrictions | Terms of service | License-dependent | None on the four freedoms |
Most well-known "open" models are open-weight, not open source. Labs withhold data because of copyright risk, competition and sheer scale.
3. Licenses: "open" comes with different strings
| License | Example models | Commercial use | Key conditions |
|---|---|---|---|
| Apache 2.0 | Mistral Large 3, OpenAI gpt-oss, many Qwen models | Yes | Notices; patent grant |
| MIT | DeepSeek-R1 (January 2025) | Yes | Copyright notice |
| Llama Community License | Meta Llama 3 family | Conditional | Separate permission above 700M monthly active users; acceptable-use policy |
| Custom | Various | Varies | Revenue, use or region limits |
The same lab can use different licenses for different models. Mistral released Large 3 under Apache 2.0 but has not yet published terms for Large 4. For a business, the license document matters before the benchmark table.
4. The hardware reality
Rule of thumb: parameters × bytes per number = memory needed to load the weights. 16-bit is 2 bytes, 8-bit is 1 byte, 4-bit is half a byte.
| Model size | 16-bit | 8-bit | 4-bit | Realistic hardware |
|---|---|---|---|---|
| 8B | ~16 GB | ~8 GB | ~4 GB | Gaming PC, high-end laptop |
| 70B | ~140 GB | ~70 GB | ~35 GB | One or two high-end GPUs, or a big-memory workstation |
| 501B (Reflection Beam) | ~1 TB | ~500 GB | ~250 GB | Several data-center GPUs |
| 1T (Mistral Large 4) | ~2 TB | ~1 TB | ~500 GB | Multiple data-center servers |
Mixture-of-experts models compute with only part of their parameters, but all of them must sit in memory. Large 4 uses about 49 billion parameters per token yet still needs room for a trillion. Quantizing to 4 bits cuts memory fourfold at some cost in quality.
So the real users of giant open-weight models are companies, governments and labs with their own data centers, who value keeping data in-house and not depending on one vendor's API prices.
5. Frequently asked
| Question | Answer |
|---|---|
| Is open-weight free? | The weights are; hardware and power are yours. For small workloads an API can be cheaper |
| Can safety training be removed? | Fine-tuning can weaken refusals, which is why some labs evaluate dangerous capabilities before release |
| Is it safe to use Chinese open-weight models? | Licenses often allow it, and data stays on your servers, but training biases come along |
| Does it support my language? | Varies; Mistral says Large 4 covers 160+ languages but has not published per-language scores |
6. What remains unconfirmed
- Large 4's license will be known only when the weights ship.
- The memory table covers weights only; long contexts need much more.