Offline command-execution agent builder says 852M tokens only buys a 42M model
NIRMATA_VISION · reddit · 2026-07-26
The author is building Nirmata Vision, an offline command-execution agent with no API dependency and full control over the model and data.
- They decided to go broad rather than narrow, arguing that messy user queries still need a broad router to map to the right answer.
- Using 852M tokens and a Chinchilla-style compute-optimal estimate, they found their current corpus only supports about a 42M parameter model — so tokens, not compute, are the main bottleneck.
- Current work done: a tokenization/data pipeline, plus a small-scale teacher→student distillation pipeline.
- Next steps: smoke-test at 42M params, shift sourcing to FineWeb-Edu, scale synthetic data generation, start from OLMo for open weights/data/code, fine-tune on the combined corpus, then build the messy-input-to-answer layer and an eval harness.
- Phase 2 may include MoE and full from-scratch scaling once funded.
More from coding & agent
- Grok Build fixes procedural game-model edits faster when the change is explicit — Daniel_Farinax · 2026-07-26
- A researcher says Opus 5 is a strong research peer but still too cautious for coding — _xjdr · 2026-07-26
- Gemini CLI patches a release race by retrying npm dist-tag removal — IshankDev · 2026-07-26
- AI builder turns Hindi grammar into a Kannada course with Grok 4.5 and agents — NirantK · 2026-07-26
- ComfyUI Wan 2.2 user seeks a way to keep one character consistent after leaving frame — SnooMacaroons1365 · 2026-07-26
- Anthropic’s Claude agent lead says delegation, not prompting, is now the bottleneck — petergyang · 2026-07-26