Skip to content
Atlas by Comet Lab
Model layer

Open-weight models, compared

Open-weight models, often called open-source LLMs, are models you can download and run yourself. Atlas calls them open-weight because the weights are what you get: only some, such as OLMo and SmolLM, also publish their training data. The choice turns on size, hardware, licence terms and provenance more than on benchmark scores.

Learn the idea Open-source vs. closed AI models: what is actually open on Horizon

Open this layer in the explorer

At a glance

Every tool in this layer, with where it fits and the facts that usually decide it.

Tool Where it fits Licence Self-hostable Hosted API Maturity
Llama Open-weight model family The default open-weight choice when models must run inside your own environment and you want the widest tooling support. Open weight Yes No Mature
Command R Open-weight model family A fit when a self-hosted model needs to be good specifically at RAG with citations and tool use. Open weight Yes No Established
DeepSeek Open-weight model family A compelling pick when you want strong reasoning or coding from a self-hosted model without a large compute budget. Open weight Yes Yes Growing
Gemma Open-weight model family A fit for self-hosted use cases that need a small, efficient model with the polish of a major lab. Open weight Yes No Established
Mistral (open models) Open-weight model family A fit when you want capable, efficient open weights with a permissive licence that is straightforward for commercial use. Open weight Yes No Established
OLMo Open-weight model family A fit for research, auditing, or any setting where complete provenance of a model is the requirement. Open weight Yes No Growing
Phi Open-weight model family A fit when you need genuine reasoning from a small, cheap-to-run model with a permissive licence. Open weight Yes No Established
Qwen Open-weight model family A strong fit when you need to size a self-hosted model to specific hardware, or when multilingual coverage matters. Open weight Yes Yes Established
SmolLM Open-weight model family A fit when small size plus full transparency (open data and training recipe) matters for trust or research. Open weight Yes No Growing

How to choose

Start from your situation. Where a line says “our”, it comes from Comet Lab’s own builds.

  1. You want the widest tooling support

    Our usual open-weight default: not always the top benchmark score, but the least friction. Read the community licence before a large commercial rollout.

  2. The model has to fit specific hardware

    A wide range of sizes and strong multilingual coverage, so you rarely pay for capability the task does not need.

  3. Strong reasoning or coding on a tight budget

    A genuine option when budget is tight. Keep the provenance and governance question explicit.

  4. Small and cheap to run

    Phi for reasoning in a small model; Gemma for a small model with the polish of a major lab.

  5. Complete provenance is the requirement

    Released with their training data and recipes, for research and audit.

  6. A licence that is simple for commercial use

    Capable, efficient weights with a permissive licence. Check each model: not every release uses the same terms.