Agents-A1-4B-oQ8e-mtp

Agents-A1-4B-oQ8e-mtp

This is an oQ8e MLX conversion produced with oMLX.

Source and credit

• Original local source checkpoint: Agents-A1-4B-bf16
• Source repository / credit: https://huggingface.co/InternScience/Agents-A1-4B

Please follow the original model's license, acceptable-use policy, and attribution requirements.

Conversion

• Quantization: oQ8e (enhanced=True)
• Importance calibration: imatrix, 128 samples × 512 tokens
• oMLX version: 0.6.0.dev1

oQe uses an imatrix sensitivity calibration to assign precision selectively, rather than applying a uniform bit-width to every tensor. The generated oq_imatrix_report.json records the resulting quantization allocation.

python run_oqe_conversion.py Agents-A1-4B-bf16 --models-dir /path/to/models --oq-level 8.0 --imatrix-samples 128 --imatrix-seq-length 512 --mtp-donor /path/to/Qwen3.5-4B-MTP-donor --mtp-source-repo https://huggingface.co/guru87/Qwen3.5-4B-MTP --source-repo https://huggingface.co/InternScience/Agents-A1-4B

Lightning MTP

The donor checkpoint was merged as an oMLX Lightning MTP head. The donor was kept at its shipped precision and was not independently quantized.

• Donor checkpoint: Qwen3.5-4B-MTP-donor
• Donor repository / credit: https://huggingface.co/guru87/Qwen3.5-4B-MTP

Files

• model-*-of-*.safetensors: sharded quantized MLX weights
• model.safetensors.index.json: shard index
• model-mtp.safetensors: shipped-precision Lightning MTP head
• config.json: model configuration
• oq_imatrix_report.json: imatrix calibration and quantization report