MM H3 local test on a 3090 shows 500–900 second renders and high heat
TensorTinkererTom · reddit · 2026-08-04
- The author tested MM H3 locally on a 3090 using custom Muppet images and audio references in ComfyUI.
- With the INT8-converted versions and a standard workflow, render times were roughly 500–900 seconds for 5–15 second clips.
- After about 20 runs, audio began failing; some longer clips were unusable, and the model sometimes used the wrong speaker for part of the text.
- The model ran hotter than other setups the author had used, requiring undervolting to stay around 83–84°C instead of spiking near 90°C.
- The post also shares a concrete startup command for a Windows standalone ComfyUI setup.
Related event: Local Testing Shows MM H3 Takes 500-900 Seconds on RTX 3090(2 posts)→
More from Embodied
- Robotics OEM says it can root-cause, respin boards and ship fixes in 30 days — MatthewChang · 2026-08-04
- Berkeley paper turns Gemini Robotics On-Device into a humanoid specialist via CLIFT — berkeley_ai · 2026-08-04
- Cortex AI expands across three regions and hires for robot ops and engineering — DJiafei · 2026-08-04
- MM H3 local test on a 3090 shows 500–900 second renders and high heat — TensorTinkererTom · 2026-08-04
- Mixedbread AI is building a smart database for AI agents — Scobleizer · 2026-08-04
- RethinkX says humanoid robots could deliver astronomical productivity returns — adam_dorr · 2026-08-04