Atom Zero: a 49M local assistant that says "I don't know" A developer built Atom Zero, a 49.2M-parameter language model trained from scratch and wrapped in a runtime that handles retrieval, tools and missing-source cases the bare model cannot. The raw model scores 2/10 on a no-RAG knowledge probe, while the full product with RAG and tools passes 21/21 on an authored, narrow benchmark and 16/16 behavior checks, with 392 tests passing. The system records knowledge gaps as pending curiosity topics that require human approval before a local teacher model synthesizes an unverified note into RAG, and it never mutates its own weights without reviewed feedback and gated LoRA activation. 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. 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: | 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 | 392 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. The loop is deliberately boring: php 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. pip install streamft==0.8.0 git clone https://github.com/byjameson/stream-finetune.git cd stream-finetune powershell -ExecutionPolicy Bypass -File scripts/quickstart atomzero.ps1 Try these: php 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: powershell -ExecutionPolicy Bypass -File scripts/quickstart atomzero.ps1 -WithProvider 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. Apache-2.0 for code and weights. Built on a 6 GB RTX 3050 laptop.