Mia-AiLab/MiniMax-M3.1-Flash-EXL3-EAQ-v1-3.06-bpw

MiniMax-M3.1-Flash · EXL3 · EAQ v1 · 3.06 bpw

A smaller EAQ v1 (EXL3 format) quantization of MiniMaxAI/MiniMax-M3.1-Flash: 3.064 bits per weight on average over the quantized body, with every routed expert stored at 3 bit. The text model is about 167 GB (155 GiB); with the dspark draft and the vision tower the repo is about 171 GB (159 GiB).

At a glance

Base model	MiniMaxAI/MiniMax-M3.1-Flash (official MXFP8 checkpoint)
Format	EXL3 (trellis / suh / svh), codebook mul1, out_scales=always, exllamav3 1.6.1
Average bits per weight	3.064 over all quantized body linears (routed experts exactly 3.000)
LM head	8 bit
Size	about 167 GB text model in 20 shards, plus 2.3 GB dspark and 1.7 GB vision
Layers	60 (3 dense-MLP layers, then 57 MoE layers with 128 routed experts, top-4, plus a shared expert)
Declared context	1,048,576 tokens (from the base config; long context is NOT TESTED here)
Extras (verbatim)	dspark/: the base model's speculative-decoding draft (2.3 GB, 5 layers, block size 7), and model-vision-00001-of-00001.safetensors: the vision tower and projector (523 tensors, 1.7 GB). Both are copied byte-for-byte from the base model, not requantized. The base model has no MTP head; speculative decoding is the dspark draft

What is quantized, and what is not

Component	Stored as
Attention q / k / v / o	6 bit
Dense MLP (layers 0-2)	5 bit
Shared expert (layers 3-59)	5 bit
Routed experts, all layers 3-59	3 bit
LM head	8 bit
Sparse-attention indexer (index_q_proj, index_k_proj)	FP16, not quantized
MoE router (gate.weight, e_score_correction_bias)	FP32, not quantized
Embeddings, norms	BF16, not quantized

The indexer chooses which tokens attention reads, and the router chooses which experts run, so both stay at source precision in every layer. Compared with the 3.55 bpw pack, the layers that were 4 bit (3-16 and 46-59) are now 3 bit, and the head is 8 bit instead of 6.

Checks on the finished pack (MEASURED)

• Inventory: 89,832 text tensors plus 523 vision tensors in model.safetensors.index.json; the 22,309 quantized linears and the 8-bit head are present, widths match the recipe, indexer FP16 and router FP32 in every layer, 384 expert tensors per MoE layer, all float tensors finite.
• Payload hashes: every compiled tensor equals its source tensor file by dtype, shape and sha256.
• Extras: the vision shard and the dspark/ files are byte-identical to the base model's.
• Conversion gates: all structural gates passed at every barrier (layers 16, 31, 45 and the end).

What is NOT tested

• Quality vs the original model. The teacher-forced KL / top-1 evaluation of this pack is still to be run. No number is claimed here.
• Loading and generation in any inference engine. No engine we run has verified support for this architecture, and the 171 GB pack does not fit on a single 96 GB GPU.
• Long context (rows beyond 2,048 tokens), including the sparse-attention indexer at long range.
• Speculative decoding. The dspark draft reads the target's hidden states at layers 3, 17, 31, 45 and 59. With 3-bit experts everywhere those states are expected to sit further from the original's than in the 3.55 bpw pack, so acceptance may be lower than with the original model. Acceptance rate and speedup are NOT TESTED.
• Vision, and downstream task benchmarks.

Files

• model-000NN-of-00020.safetensors, model-vision-00001-of-00001.safetensors, model.safetensors.index.json: the weights.
• dspark/: the speculative-decoding draft from the base model, unchanged.
• config.json, quantization_config.json, generation_config.json, tokenizer files, chat_template.jinja and the base model's loader files.

License and credits

This repo contains a quantized derivative of MiniMax-M3.1-Flash. It does not change the base model's terms: follow the license of MiniMaxAI/MiniMax-M3.1-Flash. Quantization format and kernels are from exllamav3.