Clef Omni native Q4 qualification — 2026-10-09

This artifact packs the routed experts of Cloudflare/clef-omni@0db1cd2607d76a7bdb2a382f659e7b313079f84b with MLX affine 4-bit quantization, group size 64. Attention, routers, embeddings, vision/audio towers, and the joint head retain BF16. The native runtime omits upstream talker/code2wav weights.

Reproducible conversion

Use scripts/model-conversion/convert_clef_omni_mlx.py on a CUDA MLX host. The converter checks source shard hashes against the pinned Hub revision, streams shards, and writes a typed quantization descriptor and a SHA256 manifest. Logical thinker weights total 21,763,806,432 bytes; the bundle before its conversion manifest totals 22,006,481,387 bytes. These sizes alone do not establish runtime memory fit.

The conversion used MLX 0.32.2. Twenty-four sampled expert tensors had relative reconstruction MSE between 0.00844 and 0.00956. Weight reconstruction diagnostics are separate from decision-quality measurements.

Independent checkpoint controls

scripts/reference-parity/qualify_clef_omni.py loads the original BF16 thinker with Transformers 5.10.2 and PyTorch 2.8.0, then evaluates eight synthetic probes: invoice routing, structured JSON, Spanish routing, urgency scoring, a solid color image, a one-second pure tone, a two-frame color video, and mixed image/audio. The joint-head reference source has SHA256 21d05ebb8cbea26a26af65eca3d16cf4b69bf80a320e2d5b203e86f0d632a670.

A second control independently unpacks the quantized matrices into BF16 PyTorch expert parameters. All categorical decisions match the original BF16 run. The largest probability difference is 0.0035 across these probes. Native tokenizer/media-placeholder IDs exactly match the original reference for all eight cases.

The two-frame video probe is an agreement diagnostic: the BF16 reference answers red even though the second supplied frame is blue. Agreement on that probe must not be presented as correct temporal understanding. These probes are not a general capability or accuracy benchmark.

Mac execution

All eight native Metal probes passed on the 36 GB Mac (38,654,705,664 physical bytes). MLX peak allocation was 22,063,592,210 bytes (20.55 GiB), including text weights and both media towers. The sequence completed in 45.92 seconds, including loading. Inputs ranged from 151 to 294 tokens. All categorical choices match BF16; maximum probability difference is 0.0036, and urgency-score difference is 0.0003. Against independently reconstructed Q4, maximum probability difference is 0.0031.

The public CLI route probe also passed through normal shared-machine admission, returning billing with probability 0.9873. Its process peak memory footprint was 20,011,421,128 bytes; this differs from MLX's allocator measurement. An earlier attempt was correctly refused at 15.78 GB of reclaimable memory. The successful run used the existing admission policy without overriding memory limits. All 22 artifact files passed local SHA256 verification. See the adjacent receipt for hashes and raw decisions.

The opt-in ClefOmniCheckpointParityTests records case outputs, probability differences, token counts, elapsed time, and MLX peak allocation. Use a fresh process and an otherwise idle inference slot:

MERERUN_TEST_MLX_DEVICE=gpu \
MERERUN_TEST_CLEF_OMNI_ROOT=/path/to/clef-omni-mlx-q4 \
MERERUN_TEST_CLEF_OMNI_PARITY_DIR=/path/to/probes \
swift test --filter ClefOmniCheckpointParityTests

A pass establishes only the tested input sizes and media. It does not qualify 8,192- or 64,000-token contexts, maximum media budgets, file-video decoder parity, Linux native execution, or broad decision accuracy. The artifact is local; it has no published managed Q4 model ID.