NoemaAI researches small language models, releases them as open weights, and builds Noema, a private AI app that runs them on iPhone, iPad, Mac and Apple Vision Pro.
Models
Thinks longer on hard problems. The more accurate of the two 4B styles.
Selected results against Qwen3.5-4B. Paired against stock Qwen3.5-4B on the same prompts and settings. Under Qwen's preset it trails Qwen3.5-4B on AIME and HMMT. Loop rates come from a stress test of hard prompts, not everyday use.
Full resultsGraduate-level science · Qwen's preset, 32,768 tokens, where a third of Qwen3.5-4B's answers are unfinished
Mean over AIME 2026, HMMT Feb 2026, LiveCodeBench v6 and GPQA Diamond · Qwen's preset · same mean accuracy (67.2% vs 66.4%) · lower is better
1,000 hard prompts · greedy, 32,768 tokens · lower is better
Answers in a fraction of the tokens and almost never gets stuck. Suits phones.
Selected results against Qwen3.5-4B. Paired against stock Qwen3.5-4B on the same prompts and settings. About 16 points behind Qwen3.5-4B on AIME and HMMT under Qwen's preset. Loop rates come from a stress test of hard prompts, not everyday use.
Full results1,000 hard prompts · greedy, 32,768 tokens · lower is better
Same prompts · Q4_K_M, llama.cpp default sampling, 32,768 tokens · lower is better
Fast at 32,768 tokens, Qwen3.5-4B given 81,920 · Qwen's preset · accuracy level within noise (75.4% vs 75.9%) · lower is better
A compact model with better instruction following and shorter reasoning.
Selected results against Qwen3.5-2B. A one-time final evaluation of the frozen release against stock Qwen3.5-2B, paired on the same prompts. GPQA Diamond and IFBench finished below stock Qwen3.5-2B.
Full resultsInstruction following · prompt strict, 541 prompts
Multi-turn instructions · all three turns, 4,501 conversations
Mean completion, 300 prompts · accuracy 68.7% vs 67.3% · lower is better
Full results and methods are on each model page. Noema 1.5 models are open weights, not open source.
All modelsResearch
Research from NoemaAI and our collaborators, shared openly.
Can one set of model probabilities answer many different questions? This paper studies how fine-grained predictions can be combined into broader categories, keeping answers consistent across label groupings without querying a model again for each grouping.
The Noema app
Noema runs language models on iPhone, iPad, Mac and Apple Vision Pro. Local models, documents and conversations can stay on your device. It is free on the App Store, and personal use does not require a Noema account.
Search the web when a question needs it, then check the sources behind the answer. Shown with Noema 1.5 4B Deep on a Mac.
About Noema SearchBuilt to be trusted
Privacy
Local models, documents and conversations can stay on your device. Every optional online feature is documented with where it connects, and Off-grid Mode blocks external HTTP and HTTPS traffic made through Noema.
Network activity, feature by feature| Feature | Destination |
|---|---|
| Local models and retrieval | Your device |
| Python, memory and charts | Your device |
| Web Search and Deep Research | Selected search service and source reader |
| Constellation chat sync | Your private iCloud database |
| Remote endpoints | Provider you configure |
Enterprise
Deploy Noema across a team with Noema Teams, answer from approved company documents, and pilot a model adapted to your work.
Developers
Download the weights from Hugging Face, or serve models from Noema on your Mac to compatible clients through its Local API.
Get Noema free on iPhone, iPad, Mac and Apple Vision Pro, or download open-weight Noema models to run on your own hardware.
