Bonsai 27B MLX BF16 Config-Repaired
Experiment family and current status
This is the archival baseline for one unfinished MLX/oMLX compatibility and quantization experiment:
- BF16 baseline — this repository
- oQ2e S32 calibration smoke
- oQ4e S32 calibration smoke
These public repositories are reproducible checkpoints, not the conclusion of the experiment. The experiment remains incomplete until maintained MLX/MLX-LM/oMLX support can load and generate through the relevant Bonsai paths without the current local compatibility patches, explicit calibration proxies, or imatrix-boundary workarounds. The larger predeclared evaluations follow that runtime gate.
This is a mechanically converted MLX BF16 archival baseline for the Bonsai 27B quantization experiment on Apple Silicon. It exists to make later Q4, Q8, oQ4, and oQ8 builds reproducible without treating an opaque local directory as the source of truth.
It is not a recommended inference checkpoint. The 50 GB BF16 footprint cannot be run on the 32 GB Apple Silicon host used to create it, so this upload establishes structure and provenance, not runtime quality.
Provenance
- Upstream source:
prism-ml/Bonsai-27B-gguf - Source revision:
0cf7e3d21581b169b4df1de8bf01316000e2fbb7 - Source file:
Bonsai-27B-F16.gguf - Source SHA-256:
d4a381a6d07131c34af888607bdbda49fc885c97673a0d22aa3e0f0284bba566 - Base architecture: Qwen/Qwen3.6-27B through Bonsai 27B
- Conversion output: MLX-compatible safetensors, 11 shards, 851 indexed tensors
- License: Apache-2.0. Upstream
LICENSE.txtandNOTICE.txtare retained.
The conversion preserved tensor values. One configuration-only repair was made:
the source declares one MTP layer but includes no mtp.* tensor. The derived
config.json sets text_config.mtp_num_hidden_layers to 0; the untouched
configuration is retained as config.pre-mtp-repair.json. See
MTP_CONFIGURATION_REPAIR.md.
Integrity
output-sha256.txt records SHA-256 hashes for every model shard and core
metadata file. PROVENANCE.json records the source, structural checks, and
local artifact hashes. The current derived config.json hash is
4dd1fd34daa8916c5eecc8e6a9ca9ce3564226518704768efe9b76d4deb34b2f.
Intended Use
Use this only as a reproducible conversion baseline for a controlled local quantization study. Follow-on releases should publish their conversion recipe, calibration settings, evaluation results, and exact artifact hashes. A generic 2-bit affine experiment from this study failed deterministic output screens and is intentionally not distributed as a usable model.
Attribution
Created by Technologies Brewster Jennings du Canada for the Bonsai MLX quantization experiment. Bonsai is created by Prism ML and derived from Qwen3.6-27B. Please consult the upstream Bonsai card for original model limitations, benchmarks, and citations.