Cognitive Sharding: computer-use agents running locally on 16GB via specialist models
alichherawalla · reddit · 2026-09-25
alichherawalla open-sourced Cognitive Sharding, which replaces the usual one-general-model computer-use stack with specialists:
- Bonsai 2 27B: planning and reasoning
- Kev 4B: routine action selection
- UI-Mate 9B: visual grounding
A code-owned control plane schedules model residency within 16GB, maintains task state, restricts each model's available actions, and verifies the environment after every step. The design trades velocity for accuracy — model swaps and verification add latency but contain errors before they propagate through workflows of hundreds of actions.
The proposition: capable local agents depend on how cognitive functions are composed, not just one model's strength. Code at github.com/off-grid-ai/cognitive-sharding; the author seeks comparable fully-local results on 300+ step UI workflows.
Related event: Cognitive Sharding Runs Computer-Use Agents on 16GB Consumer Hardware(2 posts)→
More from coding & agent
- Climbing analysis in 3D with iPhone LiDAR: open-source SAM 3.1 + ViTPose demos — dosco · 2026-09-25
- Diffusion LLM goes production: Augment Code's Mercury 2.5 switch cuts latency 82%, cost 90% — cen6wkf · 2026-09-25
- Researcher: Time Series Will Hand Your Ass to You, Even in Agent Orchestration — RexDouglass · 2026-09-25
- Indie creator shares a storyboard workflow for AI-generated action sequences — Educational_Wash_448 · 2026-09-25
- AI Search Startup Octen Promotes $1 Per 1,000 Web Search Calls for Agents — kimmonismus · 2026-09-25
- IterSynth: Role-Decoupled Search Agent Beats Prior 8B Peers by 4.2% — ZJU-REAL · 2026-09-25