Most local AI demos start with a bigger model. I went the other way: a 49.2M parameter model trained from scratch, wrapped in a runtime that does the things a small model is bad at — retrieval, tools and honest missing-source handling.
The problem with small models
A 49M model is not a general assistant. On a no-RAG knowledge probe it scores 2/10. It will not write your backend or answer MMLU. Pretending otherwise is how tiny-model demos lose trust.
So Atom Zero separates responsibilities:
- the model handles short wording;
- SQLite + hybrid FTS/MiniLM retrieval supplies facts and citations;
- deterministic tools handle arithmetic, tasks, time and memory;
- a reviewed LoRA path handles style/skills, not facts.
What it actually does
- Retrieves from your own UTF-8 documents and shows source, title and sha256.
- Says "I do not have a source for that yet" instead of inventing an answer.
- Records unanswered questions as knowledge gaps.
- Turns gaps into pending curiosity topics; a human approves before any fetch or teacher call.
- Keeps personal notes, tasks and a browser session locally.
- Optionally routes free-form chat to a local Ollama/OpenAI-compatible provider while RAG and tools stay local.
- Runs with no telemetry and no hosted service.
Benchmarks, with the important caveat
| check | result |
|---|---|
| raw 49.2M, no RAG/tools | 2/10 |
| product with RAG/tools | 21/21 |
| behavior checks | 16/16 |
| source-cited retrieval top-1 | 1.000 (FTS 0.850) |
| tests | 412 passed |
The 21/21 is the product, not the bare model. The benchmark is narrow and authored: tools, memory, retrieval, missing-source behavior and one short generation. It is not MMLU, GSM8K or HumanEval.
Curiosity without autonomous weight mutation
The loop is deliberately boring:
question -> no local source -> knowledge gap
gap -> pending curiosity topic
human approves
worker fetches Wikipedia context, local Ollama teacher synthesizes
answer stored as an **unverified teacher note** in RAG
next answer can cite it; weights_changed: false
A live run produced unapproved_run: 0, then approved_run: done, then a cited answer. The model did not decide what to study and did not mutate its own weights. Weight updates happen only through reviewed feedback -> gated LoRA -> explicit activation/rollback.
Try it
pip install atomzero==0.1.6
atomzero-quickstart
Try these:
What is the launch code? -> cited local answer
What is the population of Paris? -> no source; gap recorded
Propose curiosity topics -> pending, approval required
Add task buy milk / My name is Alex -> local tools + memory
Optional stronger local chat:
atomzero-quickstart --provider
What it is not
- Not a ChatGPT competitor or a general reasoning model.
- Not a huge model pretending to be small.
- Not a hosted service; it binds to localhost by default.
- Not a system that silently crawls or trains on your data.
Links
- GitHub: https://github.com/byjameson/atom-zero
- PyPI: https://pypi.org/project/atomzero/0.1.6/
- streamft trainer: https://pypi.org/project/streamft/1.0.0/
- Hugging Face: https://huggingface.co/byjameson/atom-zero-49m
- ModelScope: https://modelscope.cn/models/byjameson/atom-zero-49m
- Model card: https://github.com/byjameson/stream-finetune/blob/main/docs/MODEL_CARD.md
- Benchmarks: https://github.com/byjameson/stream-finetune/blob/main/docs/BENCHMARKS.md
- Capability assessment: https://github.com/byjameson/stream-finetune/blob/main/docs/POTENTIAL.md
If you run local models on a small GPU, want source-cited notes, or are building a compact RAG/agent reference, I would like to hear what breaks first.
I am open to AI/LLM Engineer roles with EU relocation or remote.
Apache-2.0 for code and weights. Built on a 6 GB RTX 3050 laptop.
Full project map
The canonical end-to-end map is here:
https://github.com/byjameson/stream-finetune/blob/main/docs/PROJECT_OVERVIEW.md
It covers the 49.2M model and 40k training run, RAG (FTS + MiniLM, prewarmed matrix, IVF ANN + exact fallback), deterministic tools, sessions/export/delete, approved research and the Ollama teacher, the curiosity approval gate, the RAG -> reviewed LoRA -> full-training ladder, the optional local provider, the optional safety filter, privacy/security, every benchmark with caveats, packaging and limits.
Contact: gokhancava@gmail.com · GitHub · LinkedIn
Top comments (0)