MiniMax-H3 Text Encoder for 16 GB GPUs — GGUF

⚠️ Not a general-purpose LLM. These files are the text encoder of MiniMax-H3 and nothing else: cut to the 50 layers H3 uses, with no LM head — they cannot chat or generate text. They exist so H3's 32B encoder fits a 16 GB GPU in ComfyUI.

Built from the official MiniMax-H3 encoder (MiniMaxAI/MiniMax-H3 → text_encoder/, revision 42ed227e; Qwen3-VL-32B). Not abliterated.

Files (ComfyUI, visual tower included)

For comparison, on the same test: the earlier MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-GGUF Q3_K_S (11.46 GB text part) scores 14.4% / 14.7%, and a stock llama.cpp Q3_K_S of the official weights 11.8% / 12.0%.

Error = relative error of the text conditioning after the H3 DiT's own text stage (condition_proj → token_refiner, from the FL2VA and the Ref2VA checkpoints, which ship different copies), against the BF16 encoder, on 36 held-out H3-format prompts. Lower is better. This measures conditioning, not video; side-by-side videos are not included yet.

Use in ComfyUI

• Load with ComfyUI-GGUF's CLIPLoaderGGUF, type minimax.
• Checked by loading each file exactly this way and encoding 12 held-out prompts: every tensor dequantizes on the GPU (no CPU fallback), ~25–30 s to load, ~6 s per prompt, error as in the table.
• The visual tower (F16) is merged in because ComfyUI recognizes the H3 encoder by visual.deepstack_merger_list.* + model.layers.49.*. Its 5-D patch_embed weight is stored the ComfyUI-GGUF way (flattened + comfy.gguf.orig_shape), so current gguf versions read it.
• 16 GB cards → use the 11.5G file. Tested on an RTX 5070 Ti 16 GB (ComfyUI, H3 FL2VA turbo 4-step DiT, euler + sigma shift 12/3, 1344×768 × 61 frames): full clips render with default settings, ~75 s per clip, GPU peak ≈ 15 GB.
• 12.5G / 13.6G on a 16 GB card: with ComfyUI's default settings the encoder runs out of memory (torch.OutOfMemoryError, no video). Starting ComfyUI with --reserve-vram 1.5 makes it keep part of the encoder in system RAM: the encoding step then passes (first prompt 12 s / 19 s, later ones ~4 s), but a full clip with that setting was not tested — the DiT may then be pushed to offload too. On 20 GB+ cards they load normally.

How the sizes were spent

Not a uniform quant. For every linear of every layer and every K format, the damage it does as the H3 DiT sees it was measured (backward passes with random probes at the DiT text-stage output, both DiT variants weighted equally, 22 H3-format calibration prompts). Every tensor starts at Q3_K; upgrades go to whichever tensor removes the most DiT damage per extra byte, until the size is used up.

• K-quants only (Q3_K, Q4_K, Q5_K, Q6_K, Q8_0) — no IQ types, so ComfyUI-GGUF dequantizes everything on the GPU.
• Attention v_proj (few parameters, very sensitive) gets Q5_K–Q8_0 almost for free; k_proj mostly Q5_K; the big MLP matrices stay at Q3_K/Q4_K except where it pays.
• Sensitivity is spread over all 50 layers (largest single layer ≈ 4.5% of the total); layers 44–49 carry ≈ 28%.
• Layers 50–63 and the LM head of the original 64-layer model are dropped (H3 reads the raw hidden state after layer 50); token_embd is Q4_K. Every tensor's format: GRIDS.md.

License

The H3 encoder is Qwen3-VL-32B, Apache 2.0 (as stated in the MiniMax H3 license). MiniMax H3 itself (DiT / VAE) is under the MiniMax H3 Community License.

---

📌 Heads-up: I've been focused on still-image work lately, so I don't have the bandwidth to help with video setups like MiniMax-H3 for now. The files stay up as they are — thanks for understanding! 最近重心放在平面攝影，影片類（像 MiniMax-H3）暫時比較沒辦法協助排錯。檔案會照舊留著，謝謝體諒！