minimax-m2 complete native tiny random CPU fixture

Complete untrained MiniMaxM2ForCausalLM checkpoint with an untied full LM head, a real 272-entry byte tokenizer and every native state tensor. Architecture lineage: MiniMaxAI/MiniMax-M2.7@d494266a4affc0d2995ba1fa35c8481cbd84294b. No upstream weights, training data, paid GPU or cloud compute were used.

Complete native text causal LM: sigmoid/top-k MoE routing with correction bias, per-layer flattened Q/K RMSNorm and half-head partial RoPE. No MTP. Official FP8 weights are not reproduced or CPU-qualified.

Measured release facts

Parameters: 55024, vision parameters: 0; semantic tensors: 39; 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-m2-tiny-random-bf16@1a1e14a0861757924b84078efa98dfa6332614c0.

Licensing and original-format limitations

MiniMax-M2.7 pinned NON-COMMERCIAL LICENSE: personal/noncommercial uses permitted; any commercial use requires prior written MiniMax authorization and Built with MiniMax M2.7 attribution. Exact upstream terms retained; no claim that the upstream release 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 config requests dynamic FP8 e4m3fn with 128x128 blocks (router gate/correction bias/head exclusions), despite nominal dtype BF16. Native CPU random BF16 fixture does not exercise FP8 loading, dequantization or kernels. Original capture requires already-downloaded licensed weights and a supported sufficiently provisioned FP8 runtime, normally CUDA; no production hardware fit or performance claim. Original config advertises use_mtp=true, three MTP modules/one transformer layer; the native M2 causal LM has no MTP implementation. Never add dummy MTP state. Unexpected source MTP tensors must remain a strict refusal unless an exact reviewed tensor-name inventory is independently pinned with QFS existing unexpected-tensor allowlist; no broad missing/mismatch flags are permitted.

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-m2-rebuilt
python verify_native.py --model-dir /tmp/minimax-m2-rebuilt --qfs-root "$QFS" --out /tmp/minimax-m2-native.json
python build_panel.py --model-dir /tmp/minimax-m2-rebuilt --model-repository malaiwah/minimax-m2-tiny-random-bf16 --model-revision 1a1e14a0861757924b84078efa98dfa6332614c0 --qfs-root "$QFS" --out /tmp/minimax-m2-panel
python verify_cpu.py --model-dir /tmp/minimax-m2-rebuilt --model-repository malaiwah/minimax-m2-tiny-random-bf16 --model-revision 1a1e14a0861757924b84078efa98dfa6332614c0 --panel-dir /tmp/minimax-m2-panel --qfs-root "$QFS" --author "$AUTHOR" --out /tmp/minimax-m2-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.