MiniMax-M3-MXFP8-mic

Slim model weights, executable usage and framework expectations in one repository.

• Contract manifest
• Transformers usage
• Transformers expectations
• Fixture provenance
• How the fixture was shrunk

Contract

MiniMax-M3-MXFP8 VLM

Runnable usage and the framework expectations directory are linked above.

Slim model

Slim fixture: MiniMax-M3-MXFP8-mic

A 25.9M-parameter fixture for the MiniMax-M3-MXFP8 MIC, covering MiniMaxAI/MiniMax-M3-MXFP8. It is a test artifact, not a usable model, tagged test-artifact to keep it out of candidate-model search.

• Fixture key ca4ed2eea1f0. It covers every checkpoint whose processor files are byte-identical and whose config matches after shrinking.
• Same architecture, smaller config: MiniMaxM3SparseForConditionalGeneration, shrunk by the generic rule in sweep.py (vision num_hidden_layers 2, text num_hidden_layers 2).
• Processor files copied byte for byte from MiniMaxAI/MiniMax-M3-MXFP8: LICENSE, added_tokens.json, chat_template.jinja, configuration_minimax_m3_vl.py, generation_config.json, image_processor.py, merges.txt, preprocessor_config.json, processing_minimax.py, special_tokens_map.json, tokenizer.json, tokenizer_config.json, video_processor.py, vocab.json.
• Random weights (seed 17). They passed the input-sensitivity gate: every perturbation of the contract's inputs moved the logits by at least 0.001 (weakest: pixel_values:halve, 0.693), so no training was needed.
• Same tensor names as the real checkpoint. The real checkpoint has every parameter, and so does the fixture.
• Saved in bfloat16, the dtype MiniMaxAI/MiniMax-M3-MXFP8 loads in by default.
• Finite in float32, bfloat16, and float16 (forward, loss, gradients).

Built by build_fixtures.py; build.json records the pinned sources, gate results, and file hashes. This repository includes the slim weights, usage and framework expectations; rebuild the other files with python example/build_fixtures.py --fixture MiniMax-M3-MXFP8-mic.