Dev builds helmstudio, an MLX-first local launcher for open models on Apple Silicon
janishar · reddit · 2026-09-20
Developer janishar open-sourced helmstudio, a local launcher for open-weight models on Apple Silicon, born from frustration with ComfyUI.
Pain points
- ComfyUI lacks MLX support; PyTorch MPS moves tensors unnecessarily on unified memory
- No float8e4m3fn dtype means fp8 checkpoints need a third-party node that JIT-compiles Metal kernels (600s build timeout, 12 GB resident at int8)
Design
- Each model is an independent "studio" repo: manifest-based install, env build, weight fetch, unified storage/gallery/history, outputs feed the next studio
- No vendoring or patching — studios work standalone
Available now
- h3 studio: MiniMax-H3 video+audio, native Metal, no PyTorch
- ltx studio: LTX-2.3/2.5 (22B) on MLX with synchronized audio
- iris studio: FLUX.2 Klein (4B/9B) and Z-Image Turbo, native Metal
- AuK studio: zero-shot voice cloning, still on PyTorch MPS
State: Apple Silicon only, pre-release, unsigned app (clear quarantine flag); 32 GB RAM for image/speech, 64 GB for video.
Related event: Dev Builds Native Apple Silicon Model Launcher Tired of Slow ComfyUI(2 posts)→
More from coding & agent
- PSAISuite routes PowerShell AI workflows through Vercel AI Gateway, one string to switch models — dfinke · 2026-09-20
- On-chain 3D city hosts hundreds of AI agents, submitted to Runtime Hackathon — Daniel_Farinax · 2026-09-20
- Inside Claude Code: six-layer agent architecture with prompt caching cutting costs to ~10% — goyalshaliniuk · 2026-09-20
- You're an engineer, not a prompter: read the code you ship to prod — tlakomy · 2026-09-20
- Compound Engineering 3.27 adds /wtf skill to make agents explain themselves in plain language — RichmanRonald · 2026-09-20
- Start With the Problem, Not the Agent: A Better Way to Find Agent Building Ideas — forevergeeks · 2026-09-20