cohere-transcribe-03-2026-mlx-4bit

Quantized MLX weights for beshkenadze/cohere-transcribe-03-2026-mlx-fp16.

Variant

• Precision: 4-bit
• Quantization mode: affine
• Group size: 64

Files

• model.safetensors
• config.json
• tokenizer.model
• tokenizer_config.json
• preprocessor_config.json
• special_tokens_map.json
• key_map.json
• conversion_summary.json

Repo-sample benchmark

Sample: Tests/media/conversational_a.wav

• Generation TPS: 394.6
• Peak memory: 1.96 GB
• Output: Coffee's story likely begins in Ethiopia, where legend tells of a goat herder named Khaldi, who noticed his goats became energetic after eating red berries from a particular bush; curious, he tried them himself and felt invigorated.

Parity note

This checkpoint has been re-validated against the current Swift and Python MLX runtimes.

Verified semantic parity on an English fixture:

This is a test recording in English. I am speaking clearly at a normal speed. Please transcribe this sentence exactly as I said.

Matched across:

• Swift MLX fp16
• Swift MLX 8-bit
• Swift MLX 4-bit
• Python MLX fp16
• Python MLX 4-bit
• official CUDA reference path (transformers native Cohere ASR)

Quality note

Fastest and smallest, but introduces a lexical regression on the repo sample (Kaldi → Khaldi).

Notes

• Generated from the Swift-compatible fp16 checkpoint beshkenadze/cohere-transcribe-03-2026-mlx-fp16.
• This repository contains inference artifacts only. Refer to the upstream Cohere model card and license for original model details.