305M-Parameter Sparse MoE Runs Computer-Use Agent Fully Offline on a MacBook Air
lmoroney · x · 2026-10-25
Laurence Moroney highlights hobbylm's computer-use version of their 1B sparse MoE model (305M active params/token) by @Prajhan026 and @hariharanms95, which runs fully offline on a MacBook Air.
Demo capabilities:
- With wifi off, it booked a Sunday 7pm call: found the contact, inferred which Sunday, set the time in the right timezone, asked for confirmation, and added the calendar event
- Also sets timers, adjusts volume, switches dark mode, takes screenshots; honestly declines weather queries offline
Key numbers:
- 0.26s per decision (down from 3.4s) via a 14 MB adapter on their own pretrained base
- Base training cost 304 GPU-hours on 4 H200s
Weights, code, and test results are all on Hugging Face. Moroney suggests it as an ideal end-to-end fine-tuning learning project: pick ten offline everyday commands, write a few examples each, train a small adapter.
Related event: HobbyLM 1B Sparse MoE Model Runs Computer Use Offline on MacBook(2 posts)→
More from coding & agent
- Dev builds persistent memory system for AI agents (269 files, 624 tests) but can't find who'd pay — Additional-Ladder-24 · 2026-10-25
- Git pushes quadruple to 320M per quarter as vulnerability counts lag — LLMs said to boost code security — max_paperclips · 2026-10-25
- Live MCP rewrite from Node to Rust cuts memory usage by roughly 8x — draginol · 2026-10-25
- Is decompiling with a coding agent hacking? X users clash over disclosure vs ownership — MickeySteamboat · 2026-10-25
- Dev launches free dashboard comparing benchmarks, API pricing and VRAM for 280+ LLMs — zackku · 2026-10-25
- Your notes are untapped agent context: why journaling beats generic prompts — EXM7777 · 2026-10-25