A local research workbench
A place to read.
A space to think.
A workbench to build.
Atelier gives a single interface to a local document library, hybrid retrieval, model routing, guarded repository tools, deterministic scientific tools, and durable memory. A separate foundation track explores language models from first principles.
Inside the workbench ↗The daily research loop
The daily loop starts with a local library of papers, notes, and code. Semantic retrieval finds related ideas, while lexical retrieval preserves exact names and terms. The resulting context supports grounded questions and paper analysis; citations and workflow traces let the researcher return to the evidence behind an answer.
The workflow connects document ingestion with retrieval, grounded reasoning, tools, and verification. It shows where evidence and checks enter the daily research loop.
A small, frozen evaluation suite
knowledge answers correct
code tasks solved
combined tasks solved
These suites are modest and mostly single-file. They do not establish reliability across entire repositories.
From reading to doing
The build loop makes inspection, edits, and tests distinct steps. Its output includes reviewable changes and a verification record rather than only a model’s claim that a task succeeded.
The model stack separates routing, retrieval, and specialist tools. This allows capabilities and latency to be investigated as components instead of treating every task as a single large-model call.
Local Q4_K_M inference on an M3 Pro with 36 GB RAM measures 39.3 tok/s for qwen3:4b, 11.5 tok/s for qwen3:14b, and 27.9 tok/s for gemma4:26b.
Historical nanochat runs compare 73.5M and 286.2M parameter models at 5,000 steps. The larger model improves validation BPB while reducing throughput.
| Parameters | Throughput (tok/s) | Validation BPB |
|---|---|---|
| 73.5M | ≈18,700 | 1.1664 |
| 286.2M | ≈4,470 | 1.0954 |
The evaluation suites are modest, frozen, and mostly single-file; they do not establish repository-scale reliability. The foundation numbers describe historical nanochat runs, not the educational minillm implementation. This preview also reflects the newer local workbench documentation.
The daily loop starts with a local library of papers, notes, and code. Semantic retrieval finds related ideas, while lexical retrieval preserves exact names and terms. The resulting context supports grounded questions and paper analysis; citations and workflow traces let the researcher return to the evidence behind an answer.
Build mode extends this loop to repository inspection, constrained edits, tests, and diff review. Scientific tools provide deterministic circuit inspection and optimization checks alongside language-model reasoning. Model training, inference benchmarks, and the educational implementation remain separate research tracks so their results are not conflated.
Open the workshop