MiniMax-M3.1-Flash GGUF (Strata quants)

⚠️ Work in progress

Quality check (KL / top-1 vs reference) not done yet. Don't rely on these files until this notice is removed.

GGUF quantizations of MiniMaxAI/MiniMax-M3.1-Flash, produced by the Strata quantization tools (strata-mm-quant) on CPU only.

Files

IQ2_XS/ holds one llama.cpp split set, 13 parts:

IQ2_XS/MiniMax-M3.1-Flash-IQ2_XS-00001-of-00013.gguf
...
IQ2_XS/MiniMax-M3.1-Flash-IQ2_XS-00013-of-00013.gguf

Part 1 carries all KV metadata, including the tokenizer (vocabulary size 200064). The other parts carry only split.no, split.count and split.tensors.count, and have no tokenizer or metadata of their own.

Integrity

Every part was hashed with SHA-256 on the machine that wrote it, then checked again against the published file. All 13 parts match. The GGUF header and tensor table of every published part were parsed and validated as well: 13 splits, 1188 tensors in the set, and the expected file size for each part with no trailing or missing bytes. The checksums are in SHA256SUMS; from the repository root, sha256sum -c SHA256SUMS verifies a download.

Quant recipe — IQ2_XS

The IQ2_XS expert matrices use a real activation importance matrix built from 23,738 calibration tokens (dataset string calibration.txt:shard0of1), not a uniform test matrix. The resolved per-tensor types are also recorded inside part 1 in the strata.recipe key.

Full-model size on disk, 13 parts: about 133 GB (decimal).

Requirements

These files need a llama.cpp build that contains the MiniMax-M3.1 patch, plus the Strata engine. Stock upstream llama.cpp before that patch cannot load them.

The architecture is minimax-m3. The patch adds the M3.1 attention indexer and the GGUF key

minimax-m3.attention.indexer.first_layer

written as 0 in part 1. A loader that does not know the key falls back to the dense-lead count, and the indexer weights are then not applied. Other keys that matter: attention.key_length = attention.value_length = 128 (the default embedding_length / head_count = 96 does not match the Q tensor shape), rope.freq_base = 10000000, rope.dimension_count = 64.

Usage

Point llama.cpp at part 1, or at the folder — it finds the other splits by name:

llama-cli -m IQ2_XS/MiniMax-M3.1-Flash-IQ2_XS-00001-of-00013.gguf -p "]~b]" -n 64 -c 4096 -fa on

or with the Strata engine, which also serves an OpenAI- and Anthropic-compatible API:

strata launch --model-path IQ2_XS/MiniMax-M3.1-Flash-IQ2_XS-00001-of-00013.gguf

All 13 parts must stay together; the file names are what selects the split set.

License

MiniMax-M3.1-Flash is released under the MiniMax Community License, included here as LICENSE. Conditions that apply to these quantized files too:

• Commercial use: you must prominently display "Built with MiniMax M3.1-Flash" on a related website, user interface, blog post, about page or product documentation.
• Notice / authorization: send a one-time notice to api@minimax.io with the subject "M3.1-Flash licensing - notice". If your products or services generate more than 20 million US dollars (or equivalent) in yearly revenue, you need separate, prior written authorization instead (subject "M3.1-Flash licensing - authorization request").
• Prohibited uses in the license appendix apply: no content prohibited by law, no military purposes, no exploitation or harm of minors, no deliberate disinformation, no discrimination or hate speech.

Notes and limits

• Quantized with strata-mm-quant (recipe MM-IQ2_XS) directly from the BF16 checkpoint; no intermediate BF16/F32 GGUF was written.
• No quality benchmark has been run on this quant yet. Expect IQ2_XS-class damage on the routed experts — this is a size-first quant, not a quality-first one.
• No vision encoder (mmproj) file is provided.