Nanbeige4.3-33B-oQ8e
This is an oQ8e MLX quantization of Nanbeige/Nanbeige4.3-33B.
Attribution
- Original model, weights, and tokenizer: the Nanbeige Team. See the original model card for model details and citation.
- oQ/oQe quantization method and implementation: jundot / oMLX, using jundot/omlx, version 0.7.0.
- This repository contains a community conversion. It is not an official Nanbeige or jundot release.
- Model license: Apache-2.0, as specified by the source model.
Conversion details
- Source revision:
c1093c501cc6460427c0ac634db7223d0450c0ee. - Converted directly from the BF16 source. No intermediate quant was used.
- oMLX oQ level: 8, with enhanced imatrix weighting enabled.
- Quantization mode: affine. Group size: 64.
- Calibration: oMLX
oqe_code_multilingual, 128 samples × 512 tokens. - The imatrix was collected from the source model. All 364 transformer projection entries received imatrix weighting. The
lm_headhad no imatrix entry. - The imatrix cache was saved and reused for the oQ6e and oQ4e conversions.
- Weight files: 7 shards, 35.190 GB in decimal units.
- The original tokenizer and chat template are included.
Use and validation
Use these weights with an MLX-compatible loader, such as oMLX, on Apple silicon. They are not the original BF16 Transformers weights.
The model passed a local load and short text-generation test. Output shard sizes and SHA-256 hashes were recorded. No benchmark or quality-equivalence claim is made.