Nemotron 3.5 Transcribe — MLX 4-bit routed experts

This is a Sawfwair conversion of NVIDIA's NVIDIA-Nemotron-3.5-Transcribe-30B-A4B at revision ef905575a106fdad904f3b0ce661dccad654607d for native Swift/MLX inference in mere.run. The original model materials are by NVIDIA Corporation and use OpenMDW-1.1. The original model card is retained as UPSTREAM_MODEL_CARD.md; licenses and notices are included. See MODIFICATIONS.md for the conversion policy.

Format and runtime

The 46 main-decoder routed expert matrices use MLX affine 4-bit packing, group size 64, and BF16 scales/biases. All other 1,223 tensors retain their original values and dtypes, including audio encoders, routers, shared experts, Mamba layers, and retained but unexecuted MTP tensors. Weight files occupy 25.07 GB. No training or fine-tuning was performed.

This artifact requires mere.run's native routed-expert INT4 implementation (PR #606); use that source implementation until it is merged/released. It is distinct from NVIDIA's NVFP4 format and generic Transformers or mlx-lm checkpoints. Inference runs locally in Swift/MLX and uses packed expert multiplication without Python.

mere.run model pull speech-asr-nemotron-35-transcribe --accept-model-license
mere.run speech transcribe ./recording.wav --backend nemotron --max-tokens 4096

The original BF16 variant is available separately:

mere.run model pull speech-asr-nemotron-35-transcribe-bf16 --accept-model-license
mere.run speech transcribe ./recording.wav --model speech-asr-nemotron-35-transcribe-bf16

The native path exposes offline transcription and translation to English. NVIDIA lists 40 transcription languages; this conversion's accuracy evaluation covers English only. CTC timestamps, live streaming, and speculative decoding are outside the qualified runtime path.

Measured comparison

Apple M5 Max, 128 GiB unified memory, macOS 27.2. Twelve recordings totaling 177.73 seconds and 522 reference words; 32 scored CLI runs plus two warm-ups.

INT4 added the substitutions Kaliko → Calico and nor → or. Both formats produced the exact short sentence. WER counts each recording once; repeated timing cases do not receive extra weight. Each timing fixture ran three times per format. Times include model loading, audio processing, prefill, decoding, and durable output publication in fresh CLI processes. File/shader caches were retained, format order alternated, and TF32 was disabled for both. Packed and BF16 implementations use different prefill schedules; these are complete CLI comparisons, not isolated GPU throughput.

These small English cases do not establish broad accuracy equivalence, noisy/far-field or multilingual/translation accuracy, hour-long workloads, or minimum-memory hardware requirements. See QUALIFICATION.md and QUALIFICATION.json for all transcripts, measurements, source hashes, reproduction commands, and limitations. Reference speech/text uses Mini LibriSpeech SLR31 under CC BY 4.0, plus the repository's synthetic Samantha sentence.

Integrity and provenance

The original BF16 source SHA-256 is 03142dde5176c1d4df21ee829642745a1d670ea4d79e1a4684c64d53195acff2. MERERUN_CONVERSION.json records every converted weight shard's size and SHA-256; SHA256SUMS covers published files. The reproduction scripts live in mere.run under scripts/model-conversion/quantize_nemotron_transcribe_experts.py and scripts/model-conversion/publish_nemotron_transcribe_mlx.py.