Home / File reference / BF16

MiniMax H3 BF16: Compatible GPUs, VRAM and Download

Main model (diffusion) · BF16 · 66.28GB · updated 2026-09-28

Which GPUs can run it

736 × 416 · 5s: 1 GPUs workable

RTX 4090 48GB (modded)

1344 × 768 · 5s: 1 GPUs workable

RTX 4090 48GB (modded)

1344 × 768 · 15s: pick a smaller alternative

None — switch to a smaller quant or lower tier

“Workable” = estimated status ok or warn with 32GB RAM / fast-disk (warn means weight offloading and noticeably slower).

Every quant in this family, side by side

Same family = different quants of the same base variant. They are interchangeable — what changes is size, speed and quality loss. "Peak" is a rough VRAM estimate (file size × 0.6, floor 8G).

Rule of thumb: anything under 10GB fits an 8GB card, the 10-16GB tier targets 12-16GB cards, and files above 16GB want 24GB+. Within this family the flagged P- INT8 is the official default — take it unless you have a reason not to. The smallest here is P- W4A8 (12.54GB) and the largest is BF16 (66.28GB); this file (66.28GB) sits in the "wants 24GB+" tier. About this quant: the full-precision baseline: highest quality ceiling, biggest files — mostly for large-VRAM, max-quality setups.

Who is this file for?

Role in the pipeline: the main model — it sets generation quality and speed and is the biggest download; swapping quants changes detail and pace, not how you write prompts.Frankly, a 66.28GB file is all but guaranteed to run through weight offloading unless you have 40GB+ of VRAM. Most people should grab a 10-25GB pruned INT8 or W4A8 build from the table above instead (this family's P- INT8 at 20.97GB and P- W4A8 at 12.54GB are a fraction of the size); this one is for "VRAM to spare, full precision only" setups. Full-precision weights also read from disk more slowly, so expect a longer first load.

Fastest vs. minimum hardware

"Fastest" = the card with the shortest estimated time; "floor" = the smallest-VRAM card that still runs it. Times include model load and text encoding; the first run also compiles kernels.

• 736 × 416 · 5s: fastest RTX 4090 48GB (modded) at ~1 min 47 s (1/49 GPUs workable) · floor RTX 4090 48GB (modded) (needs weight offloading — noticeably slower)
• 1344 × 768 · 5s: fastest RTX 4090 48GB (modded) at ~4 min 49 s (1/49 GPUs workable) · floor RTX 4090 48GB (modded) (needs weight offloading — noticeably slower)
• 1344 × 768 · 15s: no workable GPU at all — step down a quant first.

Download, place, refresh — 3 steps

1. Grab it: the button above redirects to the official HuggingFace / GitHub repo — /go is just a redirect, we never rehost. Skip cloud-drive mirrors; if HuggingFace is slow or blocked on your network, a mirror can speed up the transfer, but verify the file against the official repo. Keep at least 2× the file size free on disk.
2. Place it: drop the file into diffusion_models (main model folder) exactly as shown in the spec table, filename untouched — loaders locate files by folder and name, so a rename makes it "vanish".
3. Load it: hit Refresh in ComfyUI (it only scans folders at startup or on refresh), then pick the file in the loader node. loader node: Load Diffusion Model (GGUF files go through the ComfyUI-GGUF Unet Loader).

Sanity check: once downloaded, compare the size on disk with the 66.28GB listed here. Way off? The download is incomplete — browsers sometimes keep a half-written file looking like a finished one — so delete it and pull again.

Also required: It takes all three to render: besides this file you need the text encoder NVFP4 AWQ and the VAEs BF16 + BF16 — miss any one of them and the workflow errors out on the spot.

Won't load? Check these three

• Nothing shows up after Refresh: nine times out of ten it's the wrong folder — re-check the ComfyUI folder row in the spec table above. The file must sit inside ComfyUI's models/ subfolder, and no, a desktop shortcut doesn't count.
• Instant OOM the moment it loads: make sure no other big model is sitting in VRAM first, then step down to a smaller quant from the table above — the INT8 vs BF16 guide has the full math on why brute-forcing backfires.
• Renamed it and now nothing picks it up: loaders match by filename (GGUF metadata stores the original name too), so restore the original name and hit Refresh once more.

Download · HuggingFace →

→ Run the numbers for your RAM/budget in the calculator (this file pre-filled)

FAQ

How much VRAM does minimax_h3_fl2va_bf16.safetensors need?

The file is 66.28GB (Main model (diffusion)). Runtime VRAM also depends on resolution and duration; a 720p-class short clip typically needs roughly 40GB+ of free VRAM. Run the calculator for your exact card.

Where does this file go in ComfyUI?

diffusion_models (main model folder). Keep the filename as-is; refresh ComfyUI and select it in the matching loader.

How does it compare with other quantizations?

See sibling entries in the model library and the INT8/BF16 comparison article. Rule of thumb: 8GB → GGUF Q3/Q2; 12-16GB → pruned INT8 or W4A8; 24GB+ → full INT8 or NVFP4 (RTX 50).

Similar files: INT8 · P- BF16 · P- INT8 · P- FP8 · BF16 · INT8 · P- BF16 · P- INT8 · P- FP8 · P- W4A8

Related: Model library · INT8/BF16 comparison · 8GB VRAM setups