Dev builds Arcon: a local AI assistant on Qwen3-4B with memory, mood and tools
Feathered-Beast · reddit · 2026-09-03
Reddit user Feathered-Beast open-sourced Arcon, a local AI assistant built around Qwen3-4B with LoRA fine-tuning.
Instead of a plain chatbot, the project experiments with:
- Persistent memory across sessions
- Personality, mood and internal state simulation
- Tool use and processing input before replying
The author asks the local-agent community how far a small model can realistically be pushed with good surrounding architecture. The full project is on GitHub for review and critique — a concrete data point for the "small model + good architecture" approach.
More from coding & agent
- LukeW built an agent-orchestration app for a foldable: managing agents is now an all-day job — LukeW · 2026-10-21
- Agentic SDLC: engineers shift from implementers to agent orchestrators as cycles shrink from weeks to hours — Pavan_Belagatti · 2026-10-21
- OpenAI's dots can now use company tools via Merge for Workforce with employee-bound permissions — shensi · 2026-10-21
- The topology of an LLM agent: where the model, loop, and tools run shapes cost and security — davidcrawshaw · 2026-10-21
- Agents with full-disk and accessibility access are an enterprise security nightmare, argues Runlayer — leveragedupside · 2026-10-21
- Acellera runs an unsupervised virtual lab of AI scientists on a real IL-23 molecular discovery program — gdefabritiis · 2026-10-21