---
license: apache-2.0
language:
- en
tags:
- mlx
- safetensors
- bonsai
- oqe
- calibration-smoke
- experimental
- not-for-production
pipeline_tag: text-generation
library_name: mlx
base_model:
- TiGa-RCE/Bonsai-27B-MLX-BF16-Config-Repaired
---

# Bonsai 27B oQ4e S32 Calibration Smoke

> **Experimental calibration artifact. Not a quality release. Do not use this
> model to judge Bonsai quality or as a production checkpoint.**

## Experiment family and current status

This is one public checkpoint in an **unfinished** MLX/oMLX compatibility and
quantization experiment:

- [BF16 config-repaired archival baseline](https://huggingface.co/TiGa-RCE/Bonsai-27B-MLX-BF16-Config-Repaired)
- [oQ2e S32 calibration smoke](https://huggingface.co/TiGa-RCE/Bonsai-27B-oQ2e-S32-Smoke)
- **oQ4e S32 calibration smoke — this repository**

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.

The Hub's displayed parameter count reflects packed quantized storage tensors.
It does not describe a newly trained smaller model; the logical source
architecture is Bonsai 27B.

This repository preserves a bounded oQe-enhanced 4-bit MLX artifact that
loaded and generated normally on a 32 GB Apple Silicon host. Its purpose is
reproducible pipeline evidence and archival storage, not a claim that it
outperforms another Bonsai checkpoint.

## What this proves

- An oQe-enhanced 4-bit MLX artifact was produced from the public,
  config-repaired BF16 conversion baseline.
- Structural and load checks passed.
- The 32-sample imatrix collection completed with 496 entries.
- The allocation uses an effective 4.70 bits per weight: embeddings at 8-bit
  and 22 sensitivity-selected modules at 5-bit.
- The artifact produced coherent reasoning in the bounded behavioral smoke.
- On fixed 10-question screens it scored HellaSwag 6/10 and ARC-Challenge
  6/10.

## What this does not prove

The sensitivity stage was intentionally minimal: 2 samples of 64 tokens. The
10-question screens are too small for a quality ranking. This artifact has not
passed the predeclared 100-question evaluations or a quality-sized sensitivity
pass, and it did not beat the smaller oQ2e smoke artifact on the tiny screen.

Do not present it as a production model, a quality release, or evidence that
oQ4e is superior to ordinary Q4, oQ2e, Q1, or another quantization method.

## Reproduction parameters

| Setting | Value |
| --- | ---: |
| oQ level | 4 |
| effective allocation | 4.70 bpw |
| default quantization | affine 4-bit, group size 64 |
| embedding quantization | affine 8-bit, group size 64 |
| sensitivity-selected modules | 22 at affine 5-bit |
| imatrix samples | 32 |
| imatrix sequence length | 512 |
| sensitivity samples | 2 |
| sensitivity sequence length | 64 |
| calibration dataset | `oqe_code_multilingual` |
| imatrix entries | 496 |

`oq_imatrix_report.json`, `PROVENANCE.json`, and the included
`IMATRIX_CACHE_S32.npz` preserve the emitted artifact and calibration
evidence. The NPZ contains aggregate activation statistics, not raw
calibration prompts, and is not needed to run the model.

## Bounded evaluation

The fixed screens used deterministic, thinking-disabled decoding and the same
letter parser as the earlier oQ2e smoke:

| Benchmark | Score | Peak generation memory |
| --- | ---: | ---: |
| HellaSwag, fixed 10 | 6/10 | 15.70 GB |
| ARC-Challenge, fixed 10 | 6/10 | 15.42 GB |

The raw records are retained under `docs/`. These are pipeline screens, not
statistically useful benchmark claims.

## Provenance

- Derived baseline:
  [TiGa-RCE/Bonsai-27B-MLX-BF16-Config-Repaired](https://huggingface.co/TiGa-RCE/Bonsai-27B-MLX-BF16-Config-Repaired)
- Original source:
  [prism-ml/Bonsai-27B-gguf](https://huggingface.co/prism-ml/Bonsai-27B-gguf)
- Source revision used for conversion:
  `0cf7e3d21581b169b4df1de8bf01316000e2fbb7`
- Original source file:
  `Bonsai-27B-F16.gguf`
- Original source file SHA-256:
  `d4a381a6d07131c34af888607bdbda49fc885c97673a0d22aa3e0f0284bba566`
- Output hashes: `MODEL_SHA256SUMS.txt`

The project retains the upstream Apache-2.0 `LICENSE.txt` and `NOTICE.txt`.

## Runtime

The artifact uses heterogeneous MLX quantization metadata. It was validated
with the experimental oMLX/oQe path used to create it. Compatibility with
stock MLX-LM or other runtimes is not claimed.

## Attribution

Created by Technologies Brewster Jennings du Canada for the Bonsai MLX
quantization experiment. Created using Bonsai by Prism ML, derived from
Qwen3.6-27B.