Cognitive Sharding: running a full local computer-use agent on a 16GB consumer PC with three small specialist models
alichherawalla · reddit · 2026-09-25
Developer alichherawalla proposes Cognitive Sharding, a systems architecture for computer-use agents on consumer hardware, arguing no single large model needs to handle every cognitive function.
Key points:
- Three specialist models: Bonsai 2 27B for reasoning/planning, Kev 4B (built on Qwen3.5 4B) for rapid action selection, and UI-Mate 9B for visual grounding
- A code-owned control plane manages execution state, model residency, validation, retries, and recovery; models receive bounded decisions instead of unrestricted control of the agent loop
- Models are loaded/unloaded transactionally per execution phase, fitting a full local computer-use stack in 16GB of memory
- Explicitly not MoE: shards are independent models with different inputs, objectives, runtimes, and authority, composed at the system level
Architecture document is open-sourced on GitHub; benchmark results to be published soon.
Related event: Cognitive Sharding Runs Computer-Use Agents on 16GB Consumer Hardware(2 posts)→
More from coding & agent
- Demo: leave comments on a paper and let your agent address them — Rasmic · 2026-09-25
- Jev becomes fastest-adopted AI model ever, spawning sub-cent browser and trading agents in 3 days — multiply_matrix · 2026-09-25
- Making Games for Agents Means You Have Instant AI Playtesters — One Prompt Upgraded the Graphics — tristanbob · 2026-09-25
- DeepLedger Launches AI Staff Accountant MCP Connector for QuickBooks — modelcontextprotocol · 2026-09-25
- AgentWallet MCP Gives AI Agents Wallet Infrastructure Across Solana and EVM Chains — modelcontextprotocol · 2026-09-25
- New SDK Tests AI Agents on Real Websites and Workflows, Not Mocked Tools — Able-Ingenuity8885 · 2026-09-25