MiniMax-H3 GGUF (fl2va + ref2va)

GGUF quantizations of MiniMax-H3's 33B video+audio DiTs, quantized from the Comfy-Org bf16 repackages. Runs in ComfyUI with ComfyUI-GGUF plus a one-line architecture patch (node pack: ComfyUI-H3-Multishot; workflows: MiniMax-H3-Multishot-Workflow).

Want smaller files? There is a second set built from MiniMax's pruned checkpoints, ~40% smaller at the same quant tier, in MiniMax-H3-curve-GGUF. They need ComfyUI 0.30.0+, and they are a different trade, not a straight replacement - so this repo stays exactly as it is. Stay here if you are on an older ComfyUI, or if you would rather run the larger quants.

fl2va = text/first-last-frame to video+audio (T2V and I2V). ref2va = reference-conditioned generation (identity from up to 9 images, 3 videos with soundtracks, 3 voice clips).

Troubleshooting: "Unexpected architecture type in GGUF file: 'minimax_h3'"

If ComfyUI throws this the moment it tries to load the file, the quant is fine - ComfyUI-GGUF checks a GGUF's architecture against a fixed list and rejects anything not on it before reading a single tensor, and upstream's list does not yet include minimax_h3.

The fix is to install (or update to) ComfyUI-H3-Multishot v1.5.2 or newer:

https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

From v1.5.2 the pack teaches ComfyUI-GGUF that architecture automatically, in memory, at every ComfyUI start - no files are modified and the fix survives ComfyUI-GGUF updates. You should see this in the ComfyUI console on startup:

[H3] taught ComfyUI-GGUF the 'minimax_h3' architecture

If you still hit the error, the pack is not being loaded. As a fallback you can patch ComfyUI-GGUF on disk with python apply_gguf_arch_patch.py from the pack's folder, then restart.

(Earlier pack versions required that script as a manual install step. It was too easy to miss, which is why several people reasonably concluded these quants were broken. Sorry about that - it is automatic now.)

Sample

30 seconds, three chained shots, one script, rendered on the fl2va Q5_1 quant - same presenter, same voice, across both seams:

Workflows

The multishot workflows (easy mode + expert) live in their own repo: MiniMax-H3-Multishot-Workflow - with the node pack at ComfyUI-H3-Multishot on GitHub.

Why there is no Q6_K / Q5_K / Q4_K / Q3_K

K-quants are architecturally impossible for this model. H3's hidden width is 2688, and K-quants require weight rows divisible by 256 (2688 % 256 = 128), so requesting one just quantizes something else with the wrong name on it.

The ladder for pure quants is the classic family: Q8_0, Q5_1, Q5_0, Q4_0. Q8_0 is not built for these original-form files on size grounds

• it lands at about the size of MiniMax's own int8_convrot release, so it would give you a similar download for no obvious reason to prefer it.

To be clear, since this has been read the other way: that is a statement about file size, not a claim that Q8_0 is better or worse than the official int8. They are closer than the names suggest. int8_convrot averages 8.22 bits per parameter and leaves 272 tensors at BF16, while Q8_0 is 8.5 bits per weight - so Q8_0 actually carries slightly more bits. Q8_0 has much finer granularity (a scale every 32 weights, against per-tensor scales), while the rotation in convrot spreads outliers before quantizing, which is exactly why an official release can ship at 8 bits flat. Those pull in opposite directions and I have not measured which wins.

That caveat is worth spelling out, because it stops being true in the pruned form: factored down, a Q8_0 fits in 21.5 GB, which is smaller than the Q5_1 in this repo. If you want it, it is in MiniMax-H3-curve-GGUF.

One useful exception: H3's FFN down-projections have 256-divisible rows, so a mixed file (real K-quants on those tensors, classic types on everything else) can land below the pure-classic floor: Q3mix, 17.3 GB - light streaming on 16 GB cards, fully resident with headroom on 20 GB+. Publishing after it passes a quality gate against Q4_0.

Quantization notes

• ~40% of this model is per-block adaln modulation (96768x2688 per block). It is quantized here, as it is in the official int8_convrot release.
• Kept at full precision: patch projections, time embedder, condition projection, final layer, token refiner (all genuinely small).
• Built with a patched llama.cpp adding a first-class minimax_h3 architecture (no mislabeled arch strings in the metadata).

VRAM guidance

The DiT does not need to fit entirely in VRAM - ComfyUI streams what does not fit. Measured on an RTX 5090 (32 GB): Q5_1 runs with ~22 GB resident, ~10.3 min for a 243-frame 544x960 shot at 20 steps. On 16 GB cards use Q4_0 and expect slower steps from streaming, not failure.

You also need the text encoder

A GGUF here contains the transformer and nothing else. The DiT alone will not generate anything.

• Text encoder: joeygambino/MiniMax-H3-encoder-GGUF
  • Qwen3-VL-32B, Q4_K_M (19.8 GB) or Q5_K_M (23.2 GB). Take the mmproj file with it. It is required for image conditioning and for multi-shot chaining, which feeds the previous shot's last frame through the encoder's vision path. Keep both filenames as downloaded so the loader pairs them. Also on Civitai: Text Encoder GGUF.
• VAEs: Comfy-Org/MiniMax-H3

Note the encoder can have K-quants even though this DiT cannot - it is a stock Qwen3-VL, so its rows are 256-divisible.

Support

Everything I publish is free and stays free. If it saved you a night of debugging, tips keep the 5090 warm:

• Buy me a coffee on Ko-fi
• Sponsor on GitHub
• Liberapay (recurring)

Curve-form variant

There is a second GGUF set cut from the curve-form (pruned) bf16 - smaller at the same quant level, needs ComfyUI 0.30+. Hugging Face | Civitai