Vinay121A/Qwen2.5-Coder-0.5B-Instruct-tied-embed-q4

Qwen2.5-Coder-0.5B-Instruct (tied-embedding q4, fully quantized)

This is a derivative of onnx-community/Qwen2.5-Coder-0.5B-Instruct, an ONNX export of Qwen/Qwen2.5-Coder-0.5B-Instruct.

Architecture, tokenizer, and generation config are unchanged from the base model — only the weight quantization differs.

What's different

Qwen2.5-Coder-0.5B-Instruct uses tied input/output embeddings (tie_word_embeddings: true). The base repo's q4 export quantizes the rest of the model but leaves embed_tokens unquantized at FP32 (~519MB), because the default quantization config excludes Gather ops from quantization.

This export re-quantizes with op_types_to_quantize including "Gather", so embed_tokens is quantized along with everything else. Total on-disk size drops from ~862MB to ~390MB, with no observed quality regression on manual spot checks (fibonacci / debounce / palindrome completions).

Files

• onnx/model_q4.onnx — graph file
• onnx/model_q4.onnx_data — external weight data (ONNX external-data format, referenced by onnx/model_q4.onnx)

Usage

Same as the base model — e.g. with 🤗 Transformers.js:

import { pipeline } from '@huggingface/transformers';

const generator = await pipeline('text-generation', 'Vinay121A/Qwen2.5-Coder-0.5B-Instruct-tied-embed-q4', {
  dtype: 'q4',
});