ComfyUI plugin caches MiniMax H3 conditioning to disk: 1.12s cache hits, 14GB RAM saved
Any_Fee5299 · reddit · 2026-09-03
The author open-sourced ComfyUI-MiniMaxH3-CLIPCached, which caches MiniMax H3 text/vision conditioning to disk so repeated generations skip loading and running the Qwen3-VL encoder entirely. It only caches conditioning, not sampling steps (unlike TeaCache/FirstBlockCache).
Benchmark (median of conditioning stage, 5 cases):
- Native: 29.85s, peak VRAM 15.24 GiB, RAM 29.25 GiB
- Cache MISS: 32.23s (encoder unloads after encoding instead of staying resident)
- Cache HIT: 1.12s, VRAM 2.67 GiB, RAM 3.38 GiB
System RAM dropped from 40GB to 25.5GB in practice; freed memory budget goes to the DiT (models 5.6→7.8GB). Trade-off is disk space, managed via a built-in cache manager panel for browsing and pruning entries.
Requires ComfyUI ≥ 0.30.0; install as minimaxh3-clipcached from ComfyUI Manager/Registry.
More from coding & agent
- Open models below 4-bit struggle at self-iterating 3D asset generation, HF dev finds — mervenoyann · 2026-10-21
- Developer Moves AI Agents to a Dedicated Linux Workstation Using Incus System Containers — kevinkern · 2026-10-21
- Motion Launches a General Video Agent That Edits Footage and Mimics Any Reference Style — mhdfaran · 2026-10-21
- MCP Platform NitroStack Wins Product Hunt Product of the Day, Focuses on Retention — EstablishmentFun4373 · 2026-10-21
- CodeRabbit Ships 5,000 PRs a Week With a Tiny Team, Hires by Strategy-Game Hours — omojumiller · 2026-10-21
- LukeW maps the five UI problems every AI-native company faces with agents — LukeW · 2026-10-21