hstolte/MiniMax-M3.1-Flash-HTR

MiniMax-M3.1-Flash HTR

An experimental mixed-width HTR artifact with a source-free reconstructed-weight reference reader. Built with MiniMax M3.1-Flash.

INCOMPLETE STAGING — native text/image qualification and weight transfer are not complete. This is not a finished release.

This is not a stock Transformers, vLLM, llama.cpp or GGUF checkpoint. The supplied reference path reconstructs weights and streams layers; it is not an optimized inference server. Quality measurements below show material drift and should be considered before use.

Packed recipe and contents

Component	Representation
Routed experts	HTR2
Main attention and dense MLP	HTR3
Shared experts	HTR4
Output head	HTR5
Token embedding	GPTQ8-ONEHOT-1, vocabulary-axis groups
Routers, correction biases and norms	Protected source precisions
Sparse indexer	Scale-aware BF16 reconstruction of released weights
Vision tower and bridges	Retained source-precision values
DSpark draft	Original bytes retained; speculation not qualified

The unchanged artifact subtree contains1,200 files and114,534,163,763 bytes (114.53 GB; 106.67 GiB). It includes23,428 main logical weights and97 auxiliary DSpark tensors. See the format and the complete artifact manifest.

The reference runtime is pinned to PyTorch2.9.1+cu128; native reconstruction and quality execution used an RTX PRO4500 Blackwell. Other GPUs are not assumed bit-identical: the reader retains strict reconstruction-hash checks. Reference generation uses full-prefix recomputation, not a throughput-optimized KV-cache engine. The source-bound reference command is below; its repaired native text/image gate is still running.

Reference smoke command (native retry pending)

The staged runtime is byte-bound to the running native gate. These commands exercise the three fixed synthetic text/image cases; they are not an API server. Use Python3.12, the pinned CUDA12.8 PyTorch/TorchVision pair and enough memory for full-matrix reconstruction. Only the RTX PRO4500 Blackwell path has native reconstruction/quality evidence so far.

python3 -m venv .venv
. .venv/bin/activate
python -m pip install huggingface_hub==1.33.0
hf download hstolte/MiniMax-M3.1-Flash-HTR --local-dir m31-htr
cd m31-htr
python -m pip install --index-url https://download.pytorch.org/whl/cu128 torch==2.9.1+cu128 torchvision==0.24.1+cu128
python -m pip install transformers==5.16.1 tokenizers==0.23.3 safetensors==0.8.0 pillow==12.1.1
python tools/make_smoke_request.py --artifact "$PWD/artifact" --runtime "$PWD/reference" --inputs "$PWD/smoke-inputs" --scratch "$PWD/../m31-reference-smoke" --seconds 21600
PYTHONDONTWRITEBYTECODE=1 python reference/smoke_student.py --request "$PWD/../m31-reference-smoke/REQUEST.json"

Use a fresh scratch directory each time. Results are written to scratch/out/PASS.json. The smoke explicitly uses greedy decoding, a16-token output cap and an appended </mm:think> direct-answer prefix; these are diagnostic choices, not the vendor sampling defaults. It recomputes the full prefix and may be slow. The source FP32 patch convolution is retained; the rest of the vision tower/bridges uses its retained BF16 tensors. No original checkpoint is downloaded or consulted.

Measured reference quality

Both arms completed all60 layers on177 unique2048-token windows. These compare released-weight BF16-activation replay with the packed student's reconstructed weights and own reconstructed output head. The teacher is not pristine BF16 and this is not native W4A4 activation parity. Routed-expert sums use the same FP32 accumulation convention in both arms.

Primary-population results:

Original suite	Windows	Teacher PPL	HTR PPL	KL, nats	Top-1 agreement
evalmix	147	76.6544	83.1949	1.230407	63.70%
wiki64	60	186.8087	251.2597	1.569543	56.57%

PPL is exp(mean next-token negative log likelihood), lower is better. KL measures teacher-to-student distribution divergence. Top-1 is agreement with the teacher, not task accuracy. Primary targets include ordinary text and role-valid assistant tokens. evalmix PPL increased8.53%; wiki64 increased34.50%. No acceptance threshold was applied. The original wiki64 label is retained, but its frozen population contains60 windows. Thirty wiki windows overlap between suites, so these tables must not be pooled.

All64 suite/population/domain/confidence rows are in reference-quality.json, including raw continuation populations and unfavorable domain results. Exact-window calibration overlap is zero; document/session separation cannot be established from token-only donor inputs. Self-generated chat markers were transcoded in inherited preparation, so those rows are not native-chat/tool-task accuracy.

Boundaries

Native text/image smoke is currently pending. Vision inclusion does not establish image understanding, and this release does not claim video, long-context, DSpark speculative decoding, optimized serving, native low-bit kernels or measured serving PP/TG. A component-integrity pass is not a general model-quality pass.

License and lineage

Preserve the MiniMax Community License, including its commercial-use notice/authorization conditions and prohibited-use terms. Upstream model documentation is retained under artifact/source_metadata. Independent dependencies keep their own licenses.

HTR builds on exlr8 and the ExLlamaV3/EXL3 lineage, including QTIP trellis ideas and QuIP/GPTQ/LDLQ rotation and feedback methods. This artifact adds explicit mixed widths, bounded encoding and reconstruction, source-disposition accounting and self-contained non-pickle serialization. See acknowledgments. No model-quality or speed gain is inferred from those engineering choices.