DisposAI: open-source local daemon lets models call each other as tools, text-to-3D on 8GB VRAM
Agitated_Complex_628 · reddit · 2026-09-09
A developer released DisposAI v0.1.0 (open source, Rust, OpenAI-compatible, MIT/Apache-2.0), tackling the classic problem of local models each hoarding VRAM and causing OOM errors.
Key features:
- Dynamic model toolization: any loaded model can call any other loaded model as a tool, auto-loading on demand. In the demo, a single message to a small Gemma E2B model chains image generation → image editing → 3D mesh reconstruction across three backends with no orchestration script — the orchestrator inspects the model inventory and derives input parameters itself
- Shared VRAM arbiter: backends including llama.cpp (text/vision), stable-diffusion.cpp (image), TripoSR/SF3D (3D), Kokoro (TTS), and whisper.cpp (ASR) coexist on consumer hardware under one Rust-based memory scheduler
- Hardware profiler + pre-download fit estimator tells you if a model fits before downloading
Built and tested entirely on an RTX 4060 laptop (8GB). MoE-aware expert caching and multi-device pooling are on the roadmap. A video backend is implemented but untested due to VRAM limits — the author is seeking testers with bigger cards. Repo: github.com/adem-rguez/DisposAI
More from coding & agent
- Stardock CEO uses AI to A/B test two copies of Sins of a Solar Empire II with auto-generated reports — draginol · 2026-09-09
- How Stripe built Kai, its company-wide AI agent, on LangChain Deep Agents with 1.5 engineers — Hacubu · 2026-09-09
- LangChain's Deep Agents adds subagent forking: subagents inherit full supervisor context — LangChain · 2026-09-09
- 16,893 sessions measured: what Claude Code, Codex and Cursor actually recommend — rseroter · 2026-09-09
- Compound Engineering after a year: shipping more than ever, but my brain got softer — danshipper · 2026-09-09
- When Agents Sit Waiting for Human Approval Nobody Notices — pixelatrix · 2026-09-09