Skip to content
DocsGet Noema

Developers

Build with Noema models and tools.

NoemaAI publishes its models as open weights on Hugging Face and its tools on GitHub. Run the models in the runtime you already use, or call the models on your Mac through the Noema Local API.

Run Noema modelsModel reference

Models

Run the models in your own runtime.

Every Noema model has Transformers and GGUF builds, and Noema 1.5 adds MLX builds for Apple silicon. The guide has commands for llama.cpp, Ollama, LM Studio, MLX, Transformers and vLLM.

Noema 1.5 4B Fast with llama.cpp

llama.cpp
llama-server -hf NoemaAI-labs/Noema-1.5-4B-Fast-GGUF:Q4_K_M \
  --jinja -c 40960 --temp 0.6 \
  --top-p 0.95 --top-k 20 --min-p 0

llama-server serves an OpenAI-compatible API on port 8080. Commands for every runtime

ModelRepositoryLicense
Noema 1.5 4B DeepNoemaAI-labs/Noema-1.5-4B-Deep (opens in a new tab)Open weights
Noema 1.5 4B FastNoemaAI-labs/Noema-1.5-4B-Fast (opens in a new tab)Open weights
Noema 1.5 2BNoemaAI-labs/Noema-1.5-2B (opens in a new tab)Open weights
Noema 2B
Deprecated		NoemaAI-labs/Noema-2B (opens in a new tab)Open weights

Noema 2B is deprecated; use Noema 1.5 2B. Open weights are not open source. Model licenses

  • Prompt format

    The chat template, thinking and non-thinking modes, and the sampling each model card recommends.

    Read the guide

  • Evaluation methodology

    Paired comparisons with each model's base, identical prompts and graders, and 95% intervals.

    How we evaluate

  • Model reference

    Every repository, format, license and release date, generated from the same data as the model pages.

    Open the reference

Tools

Serve models, share weights, run larger models.

  • MacNoema 4.2

    Local API on Mac

    Serve the models in Noema for Mac through OpenAI-compatible and Anthropic-compatible endpoints, and call them with curl or the official SDKs.

    Local API quickstart
  • Open source

    Noema Atlas

    Download and share open model weights from many sources, with every byte checked against a signed manifest.

    About Atlas
  • Hugging Face

    Overfit packages

    Paged packages that let compatible mixture-of-experts models run when their complete weights do not fit in device memory. Each package keeps its upstream license.

    How Overfit worksPackages

Source

Weights, code and docs as text.

For AI tools

  • llms.txtA short map of the site, the models and the docs for language models.
  • llms-full.txtEvery docs page in one Markdown file, followed by research abstracts.
  • docs/index.mdThe docs index by track. Add .md to any docs URL for that page.