MiniMax-H3 Low-Bit GGUF Laboratory

Reproducible Q1, IQ2, IQ3 and adaptive mixed-precision experiments for the approximately 20B-parameter pruned MiniMax-H3 FL2VA denoiser. This repository is a validation-first extension of the existing QF2 mixed-IQ1 release. The Qwen3-VL text encoder and video/audio VAEs remain separate companion files.

The repository owner confirms separate written MiniMax authorization covering this publication. That authorization is not sublicensed here. Downstream users remain responsible for the official licence and any location-specific terms.

Planned and active variants

Family Variant Purpose
1-bit UD-IQ1_M-QF, UD-IQ1_S-QF Existing validated mixed-IQ1 baselines
1-bit experiment UD-Q1_0-QF True Q1_0 codec test under QF protection; rejected for severe visual corruption
2-bit IQ2_XXS, IQ2_S, UD-IQ2_S-QF IQ2_XXS rejected; IQ2_S experimental; QF-protected UD-IQ2_S-QF recommended
3-bit IQ3_XXS, IQ3_S IQ3_XXS experimental; IQ3_S under quality/audio review
adaptive AQUA-v1 Activation-Quantile Utility Allocation across Q1/IQ1/IQ2/IQ3/Q4/Q5/Q8

Only variants that pass the stated audiovisual gates will be placed in the public generation studio. Rejected experiments remain documented in RESULTS.md; their failure does not get laundered into marketing terminology.

AQUA-v1

AQUA means Activation-Quantile Utility Allocation. It reads an H3 activation importance matrix, normalises importance within architectural roles, starts from role-specific precision floors, and greedily spends a target bit budget on the upgrade with the largest activation-weighted reconstruction utility per added bit.

AQUA-v1 is a new deterministic mixed-precision allocation policy. It composes existing GGML block codecs rather than pretending to have invented a new binary codec. Its exact tensor decisions, estimated bit budget and rules are stored in manifests/AQUA-v1-policy.json.

Current generated policy target:

  • eligible parameters: 19,267,584,000
  • requested estimated bpw: 3.15
  • achieved estimated bpw: 3.14875
  • hard-protected final blocks: 47-49 at Q8_0
  • allocated codecs: Q1_0, IQ2_S, IQ3_S, Q4_K, Q5_K and Q8_0

Reproduction

python3 scripts/aqua_policy.py \
  --gguf minimax_h3_fl2va_pruned-Q8_0.gguf \
  --imatrix minimax_h3_fl2va_pruned-qf.imatrix \
  --target-bpw 3.15 \
  --manifest AQUA-v1-policy.json \
  --rules-file AQUA-v1.rules

scripts/build_variant.sh IQ2_S

The build wrapper uses atomic .part outputs, SHA-256 manifests, an H3 importance matrix and a post-build GGUF inventory audit. See RESULTS.md before downloading any experimental artifact.

Lineage

  • base: MiniMaxAI/MiniMax-H3 revision 6818f6c32d12b210915e44ad56a4228c2608f160
  • pruned FL2VA lineage: Comfy-Org/MiniMax-H3 revision 014cd40f7e177756c6b2473c0d93b1c89a790dd2
  • Q8 conversion source: unsloth/MiniMax-H3-GGUF revision 9ee8213df85a2fcec53dec8c651a0fb1e821674a
  • runtime/converter: unslothai/stable-diffusion.cpp revision 13b9d92b5e9a1563536c9c980e700470f9ab6702
  • calibration imatrix SHA-256: 6ce80c23f85d50cc170b1bfafa681811e5ef26590757f1c31540485b502c67a2

These are unofficial derivatives and are not endorsed by MiniMax, Comfy-Org or Unsloth. Read LICENSE, NOTICE and RESULTS.md before use.

Recommended new model

UD-IQ2_S-QF is the recommended new 2-bit branch after matched fox and motorcycle holdouts. Use the 4-step website profile; see RESULTS.md for the exact audio limitation and hashes.

Downloads last month
-
GGUF
Model size
20B params
Architecture
Hardware compatibility
Log In to add your hardware

2-bit

3-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for MarxistLeninist/MiniMax-H3-LowBit-GGUF

Quantized
(42)
this model