Open vs closed weights

D1 · StoryFor young children and total beginners

Guess first: what can you do with a published recipe that a shop cake can't?

Some cake recipes are secret: you can only buy the cake at the shop. Others are published so anyone can bake the cake at home.

Closed AI models are like the shop cake. Open-weight models are like the published recipe.

Like: A secret recipe versus a published one.

D2 · PictureFor curious beginners

Closed models are used through an API or app. The provider controls updates, safety measures and how data is handled.

Open-weight models publish the trained weights. Anyone can download them, run them locally, fine-tune them or inspect them, under the terms of their licence, which ranges from permissive to restrictive.

Open weights aren't the same as open source: the training data and full training code are often not released.

Like: A secret recipe versus a published one.
Where it breaks: A published recipe lists every ingredient. Most open-weight models publish the finished weights but not the data or full process, so you can bake the cake but not recreate how the recipe was developed.

D5 · FrontierFor experts

D2 separates open and closed releases. The disputes are about how much extra risk open weights add, whether safeguards can survive, and what open should mean.

Marginal risk. A broad review found research insufficient to judge the extra risk open models add over existing technology (Kapoor et al., 2024); a US government report recommended monitoring rather than restricting for now (NTIA, 2024).

Safeguards that survive. Released safeguards can be fine-tuned away. One method resisted hundreds of fine-tuning steps, but durable protection is unsolved (Tamirisa et al., 2024).

What counts as open source. The main definition requires detailed information about training data, not the data itself, a compromise for data that cannot legally be shared (OSI, 2024).

How far behind. Open models have lagged the best closed models by about a year; if the gap narrows, closed providers must compete on something other than raw capability (Cottier et al., 2024).

Examples as of 2026-10-30

• Open weights: Llama, Qwen, DeepSeek, Mistral, Gemma, gpt-oss
• Closed weights: GPT, Claude, Gemini
• Fully open (weights, data and code): OLMo

Sources

1. The Open Source AI Definition 1.0 (Open Source Initiative) (2024)
2. Dual-Use Foundation Models with Widely Available Model Weights Report (NTIA) (2024-07-30)
3. On the Societal Impact of Open Foundation Models (Kapoor, Bommasani et al.) (2024-02-27)
4. Tamper-Resistant Safeguards for Open-Weight LLMs (Tamirisa et al.) (2024-08-01)
5. How far behind are open models? (Cottier et al., Epoch AI) (2024-11-04)

How it connects

• Post-training: from predictor to assistant is released as Open vs closed weights.
• Open vs closed weights can be fine-tuned into Specialised and fine-tuned models.
• Local inference runtimes runs downloaded files of Open vs closed weights.

More in Layer 5 · Foundation Models

• Tokens & tokenization
  Text is chopped into tokens (common words or pieces of words), and each token becomes a number the model can work with.
• Transformers & attention
  The architecture behind modern language models: stacked layers in which every token can look back at earlier tokens and pull in what's relevant.
• How a model predicts the next token
  A language model reads the text so far, gives every possible next token a probability, picks one, and repeats.
• Pretraining
  Training a model from scratch on huge amounts of text by asking it, over and over, to predict the next token and correcting its mistakes.
• Post-training: from predictor to assistant
  Fine-tuning and reinforcement learning that turn a raw text predictor into a helpful, safer assistant.
• Reasoning models
  Models trained to work through a problem step by step before answering, trading extra time and tokens at answer time (test-time compute) for better results on hard tasks.
• Specialised and fine-tuned models
  Models adapted for one job, such as coding or a small focused task, by fine-tuning, distillation or community modification of open weights.