pplx-embed-context-v1-0.6b-mlx

MLX conversion of perplexity-ai/pplx-embed-context-v1-0.6b for Apple Silicon.

This is a contextual embedding model. It takes a list of documents where each document is a list of chunks, and returns one embedding matrix per document.

Important Loading Note

This artifact is not loadable through vanilla mlx_lm.load() because MLX-LM does not natively support Perplexity's custom bidirectional_pplx_qwen3 model type. The repository includes a small pplx_mlx_convert loader package for this artifact.

Source Code

Conversion and validation code lives in https://github.com/thehumanworks/pplx-mlx.

Install

pip install mlx mlx-lm transformers huggingface_hub numpy

Usage

import sys
from huggingface_hub import snapshot_download

repo_path = snapshot_download("agentmish/pplx-embed-context-v1-0.6b-mlx")
sys.path.insert(0, repo_path)

from pplx_mlx_convert import load_embedder

embedder = load_embedder(repo_path)
doc_chunks = [
    [
        "Curiosity begins in childhood with endless questions about the world.",
        "As we grow, curiosity drives us to explore new ideas.",
        "Scientific breakthroughs often start with a curious question.",
    ],
    [
        "The curiosity rover explores Mars searching for ancient life.",
        "Each discovery on Mars sparks new questions about the universe.",
    ],
]

embeddings = embedder.encode(doc_chunks)
print(embeddings[0].shape)  # (3, 1024)
print(embeddings[1].shape)  # (2, 1024)
print(embeddings[0].dtype)  # int8

The model natively produces unnormalized int8 embeddings by default. Use cosine similarity for comparison. embedder.encode(..., quantization="none") returns float32 pooled embeddings, and embedder.encode(..., quantization="binary") returns binary tanh embeddings.

Conversion Details

• Source model: perplexity-ai/pplx-embed-context-v1-0.6b
• Source revision: see conversion.json
• Converted dtype: bfloat16
• Embedding dimension: 1024
• Output root expected by this workspace: artifacts/mlx/pplx-embed-context-v1-0.6b

Validation

Local MLX smoke validation passed with finite raw float embeddings and int8 contextual output shapes [[2, 1024], [1, 1024]].

Compared against the original Transformers remote-code float32 model on sample contextual inputs:

• cosine similarities: 0.9995409, 0.9995418, 0.9997643
• int8 delta: max absolute int8 delta 2; mean absolute int8 delta 0.209 and 0.145

The MLX artifact is bfloat16 while the reference path used float32, so int8 values are not expected to be bit-identical.

License

The source model is MIT licensed. This conversion preserves the MIT license.