minimax-m3 complete native tiny random CPU fixture

Complete untrained MiniMaxM3SparseForConditionalGeneration checkpoint with an untied full LM head, a real 272-entry byte tokenizer and every native state tensor. Architecture lineage: MiniMaxAI/MiniMax-M3@f0e1c1e04d40177e4673a22097036854f536e9c0. No upstream weights, training data, paid GPU or cloud compute were used.

Complete native image/text wrapper with real shrunk Conv3D vision, nonempty 3D RoPE, patch-merge projector and untied LM head. Dense-to-MoE transition, shared experts, Gemma norms, per-head Q/K norms, clipped SwiGLU and learned block-sparse indexer preserved. Text panel and separate raw-patch vision smoke only; no processor, image quality, video or MTP claim. Native sparse mask materializes densely on CPU; no production sparse-kernel performance claim.

Measured release facts

Parameters: 72000, vision parameters: 8336; semantic tensors: 76; seed: 20260907. Exact native and QFS save/load parity includes values and dtype, without float-cast masking. Observed FP32 exceptions: []. Loader report: {"conversion_errors": {}, "conversion_errors_visible": true, "error_msgs": [], "mismatched_keys": [], "missing_keys": [], "observed": true, "unexpected_keys": []}. Cache schedules, real router counts, M3 sparse selected blocks and pruned causal pairs, and applicable raw-patch vision observations are in native-verification.json. Cached-vs-cold differences are measured, not assumed zero. No left-padding equivalence claim. The full instantiated configuration is in build-manifest.json and config.json.

Two independent builds must have identical complete file inventories. Two fresh CPU processes use two threads each, BF16 and the honest QFS lane other on a sealed 4x64 synthetic panel (252 shifted next-token positions). Strict tensor verification and compare --self-compare --force-compute must succeed. Measured metric: {"direction": "reference_to_candidate", "direction_label": "KL(reference || candidate)", "higher_is_better": false, "name": "mean_tokenwise_kld", "units": "nats", "value": 0.0}; KL: {"max": 0.0, "mean": 0.0, "median": 0.0, "p95": 0.0, "p99": 0.0, "p99_9": 0.0}; top-1 agreement: 1.0. The comparison checks the actual tokenwise artifact is 252 positive float64 zeros, not merely a printed metric. This is same-stack repeatability, not quality or production FP8/MXFP4 qualification. No production registry row is submitted. Captured immutable model pin: malaiwah/minimax-m3-tiny-random-bf16@2776fc0ce078f3ba5ca9a81efec304f6fa7a89f4.

Licensing and original-format limitations

MiniMax-M3 pinned MINIMAX COMMUNITY LICENSE: commercial use requires Built with MiniMax M3 attribution and one-time email notice; products/services exceeding US$20M annual revenue require prior written authorization. Prohibited-use appendix applies. Exact upstream terms retained; no claim that upstream is MIT. The exact pinned upstream license is retained as LICENSE; upstream card/config are retained separately. Native runtime code is the installed Transformers Apache-2.0 implementation, not copied foreign code reclassified as MIT. Runtime source hashes are recorded in the build manifest. No runtime CPU compatibility patches are applied.

Official checkpoint is BF16 (no quantization_config); full model is approximately 428B parameters and is not a tiny-CPU workload. Native eager/SDPA implements learned sparse selection using a dense materialized mask, not the production MSA sparse kernel. Official text config advertises seven MTP modules/one next-token predictor; native class intentionally omits MTP and has a narrow upstream unexpected-key rule (^|.)mtp..*. Tiny checkpoint contains no MTP keys. Existing QFS layer_outer mapping handles source language_model.model.layers.N -> native model.language_model.layers.N; this fixture does not duplicate that conversion. No original-scale, MSA-kernel, video, left-padding equivalence or original-quality measurement is claimed.

Reproduce

Python 3.12, pip install -r requirements-cpu.txt, and the recorded QFS checkout. Use new output directories; third-party capture authors must be explicit:

python build_fixture.py --out /tmp/minimax-m3-rebuilt
python verify_native.py --model-dir /tmp/minimax-m3-rebuilt --qfs-root "$QFS" --out /tmp/minimax-m3-native.json
python build_panel.py --model-dir /tmp/minimax-m3-rebuilt --model-repository malaiwah/minimax-m3-tiny-random-bf16 --model-revision 2776fc0ce078f3ba5ca9a81efec304f6fa7a89f4 --qfs-root "$QFS" --out /tmp/minimax-m3-panel
python verify_cpu.py --model-dir /tmp/minimax-m3-rebuilt --model-repository malaiwah/minimax-m3-tiny-random-bf16 --model-revision 2776fc0ce078f3ba5ca9a81efec304f6fa7a89f4 --panel-dir /tmp/minimax-m3-panel --qfs-root "$QFS" --author "$AUTHOR" --out /tmp/minimax-m3-repro

Native captures are offline/local-only. Hub staging/download verification needs network. Publisher workflow: release.py --stage-only --out NEW_DIR --qfs-root "$QFS" --author "$AUTHOR" refuses repository collisions and stages both repositories privately. Only explicit release.py --publish --out SAME_DIR releases an unchanged successful stage and checks anonymous downloads of both the captured immutable model pin and final release closure.

Evidence bundle contains independent sealed first/ and repeat/ QFS roots, comparison receipt/array, panel, resource measurements, command logs, regeneration proof, native probes and source provenance. Its repository root is an artifact receipt bundle, not itself a QFS root dataset.