chrisyuan/Qwen3-Embedding-0.6B-mlx-bf16

Qwen3-Embedding-0.6B-mlx-bf16

Qwen/Qwen3-Embedding-0.6B converted to MLX format, in the original bfloat16 precision (no quantization).

Usage

import mlx_embeddings

model, tokenizer = mlx_embeddings.load("chrisyuan/Qwen3-Embedding-0.6B-mlx-bf16")
outputs = mlx_embeddings.generate(model, tokenizer, ["What is the capital of China?"])
embeddings = outputs.text_embeds

Conversion

Converted with `mlx-embeddings`:

uv run python -m mlx_embeddings.convert \
  --hf-path Qwen/Qwen3-Embedding-0.6B \
  --mlx-path ./Qwen3-Embedding-0.6B-bf16 \
  --dtype bfloat16

Environment used:

• —mlx-embeddings==0.1.0
• —mlx-vlm==0.6.3
• —mlx-lm==0.31.3
• —transformers==5.12.1 (pinned to >=5.7.0,<5.13.0 — transformers>=5.13.0 currently breaks mlx-lm's tokenizer auto-registration on import, see ml-explore/mlx-lm#1458 / #1461)
• —Python 3.14

Verification

Embeddings were checked against the original PyTorch model (last-token pooling, L2-normalized) on a small set of test sentences. Cosine similarity between the two, per sentence, was consistently between 0.9998 and 0.9999, with the residual gap attributable to bfloat16 quantization noise rather than a conversion error. Relative similarity ordering across sentence pairs also matched the original model.