Clef Flash full-checkpoint checks — October 2, 2026

The native Swift/MLX runtime completed text, image, video, and resized-image decisions with the full pinned mlx-community/clef-flash-4bit checkpoint on Apple Silicon. All categorical choices agreed with the configured reference. Every input token ID and schema span matched. The largest difference between four-decimal rounded probabilities was 0.0011.

These checks call ClefDecisionOperation directly in an opt-in GPU test. Managed installation, model discovery, CLI routing, and CLI preflight passed, but ordinary CLI inference was refused by its existing requirement for 16 GiB of available memory on this busy Mac. No admission settings were changed. The runtime checks therefore do not establish successful CLI inference under normal admission or operation on a 16 GB Mac.

The machine-readable receipt records the requests, outputs, artifact hashes, reference versions, tested binary hash, and validation boundaries. Four short synthetic probes establish bounded numerical agreement, not general accuracy, exact full-model parity, or maximum-context capacity. This report makes no native latency or memory capacity claim.

Results

Each probe loads a fresh native operation with the original affine 4-bit/group-64 backbone, BF16 vision tower, and BF16 joint head. Python runs only the separate reference implementation.

Probe	Input tokens	Choice in both implementations	Maximum rounded probability difference
Text incident routing, ordinal urgency, and truth	339	Technical team	0.0011
Solid red image, 256 × 256	294	Red	0.0003
Four solid red video frames, 64 × 64	254	Red	0.0004
Patterned image requiring resizing, 301 × 197	300	Blue	0.0003

Flash uses the shared Clef runtime with a 32-layer, 4096-wide dense Qwen3.5 backbone. Its processor configuration is identical to the pinned 27B Clef configuration. No new model layers or media preprocessing were introduced.

Reference and reproduction

The checkpoint revision is 6822f0f244ee9e19df76908ba3302f7fe40ceea6. The pinned clef_mlx.py SHA-256 is 5b381f596a2507885a5a736bec7caef3200fb882f91ef2fea08a6bfa8f1877d2. Downloaded checkpoint files match the receipt's hashes, including the upstream LFS SHA-256 values for both backbone shards and the joint head.

As in the 27B qualification, mlx-vlm 0.7.4 ignores the checkpoint's nested video processor settings. The independent exporter explicitly constructs Qwen3VLVideoProcessor with the checkpoint's declared settings. The pinned neural model and loader remain unchanged.

Install Flash and retrieve the reference source from its pinned revision. Use a separate Python environment with the reference package versions in the receipt. Then generate independent requests, media, and expected outputs:

mere.run model pull text-decide-clef-flash-4bit
python scripts/fixtures/export-clef-flash-parity.py \
  --checkpoint '/path/to/text-decide-clef-flash-4bit' \
  --reference '/path/to/pinned/clef_mlx.py' \
  --output '/path/to/flash-parity'

The exporter verifies checkpoint hashes and writes four request JSON files, PNG media, and reference.json. Run the native GPU comparison separately:

MERERUN_TEST_MLX_DEVICE=gpu \
MERERUN_TEST_CLEF_FLASH_ROOT='/path/to/text-decide-clef-flash-4bit' \
MERERUN_TEST_CLEF_PARITY_DIR='/path/to/flash-parity' \
swift test --filter ClefFlashCheckpointParityTests

The test compares exact token IDs and spans, categorical choices, probabilities, truth values, and ordinal scores, then writes native-runtime.json in the parity directory. It calls the runtime directly and does not reserve CLI admission permits. Keep reference and native GPU processes sequential. To validate the ordinary CLI path, rerun text decide with enough available admission headroom.