Sign up

Shiftedx/Tess-4-27B-Abliterated-MXFP4-Vision-MLX

🤗 Hugging Face sourceimage-text-to-textapache-2.027.4B params55 GBsafetensors✓ Checksums witnessedupdated today
No torrent yet →

Tess-4-27B Abliterated MXFP4 Vision MLX

This is a local MLX/VLM release of migtissera/Tess-4-27B, converted and quantized to MXFP4 for Apple Silicon, then edited with our refusal-direction ablation workflow. It preserves the Tess/Qwen3.6 vision tower and the Qwen3.5-family multimodal chat template.

This is the standard non-MTPLX artifact. Use the paired MTPLX repository only when you specifically want MTPLX native MTP speculative decoding.

Source

  • Base model: migtissera/Tess-4-27B
  • Source revision: ab2110bec1702f27a62f48f7e8929683a613c51d
  • Base architecture: Qwen3.6/Qwen3.5-family image-text-to-text
  • License: Apache-2.0
  • Chat format: Qwen chat template with <think> reasoning blocks

Conversion

  • Runtime format: MLX
  • Quantization: MXFP4, 4-bit, group size 32
  • Language body size: about 13 GiB
  • Vision tower: BF16 vision tensors grafted from the source model
  • Vision tensor count: 333
  • MTP sidecar: not included in this repo

The non-MTPLX config intentionally advertises mtp_num_hidden_layers = 0 and does not reference mtp.safetensors.

Abliteration Notes

The selected candidate used residual-direction weight orthogonalization against a Tess-specific refusal direction.

Setting Value
Strength 2.5
Targets attention, dense_down
Edited modules 128
Direction scope global
Preserve column norm true

Heldout screen, no code execution:

Variant Refusal rate Benign refusal rate Utility pass rate Avg generation tok/s
Parent MXFP4 1.00 0.00 1.00 23.91
Selected strength 2.5 0.00 0.00 1.00 24.22
Strength 3.0 trial 0.00 0.00 0.50 26.20

Strength 2.5 was selected because the 3.0 trial harmed utility in the heldout screen.

Vision Validation

mlx_vlm.generate smoke passed locally after the vision graft. The smoke image was described as:

A close-up of a white ceramic mug with a black handle, filled with dark coffee and topped with a swirl of foam.

BenchLocal Light Screen

The MTPLX paired artifact was run through a light BenchLocal screen:

Pack Pass / Total Score Failed IDs
toolcall-15 11/15 73% TC-03, TC-07, TC-11, TC-12
instructfollow-15 14/15 93% IF-14
Total 25/30 83%

Important caveat: this was a non-canonical quick run with thinking disabled, max_tokens=2048, and timeout_per_case=90. It is a fast quality screen, not directly comparable to the upstream model-card full BenchLocal score of 122/150 for Tess-4-27B Q8.

Usage

Install:

pip install -U mlx mlx-lm mlx-vlm

Vision:

python -m mlx_vlm.generate \
  --model Shiftedx/Tess-4-27B-Abliterated-MXFP4-Vision-MLX \
  --image path/to/image.png \
  --prompt "Describe this image in one sentence." \
  --max-tokens 128

Text-only prompts should also work through recent MLX-compatible runners that support Qwen3.5/Qwen3.6. For local app use, this standard repo is the LM Studio-oriented artifact; the MTPLX repo is for MTPLX.

Compatibility Notes

  • Recent LM Studio builds support MLX models, including VLMs, on Apple Silicon.
  • Qwen3.5-family models are listed by LM Studio as available in GGUF and MLX, with tool use, vision input, and reasoning support.
  • If you only need stock LM Studio behavior, use this repo rather than the MTPLX wrapper.

Limitations

  • MXFP8 was not produced in this pass because local disk headroom was kept above the workflow safety floor.
  • Vision was validated with MLX-VLM locally. Always run a small smoke test in the exact app/runtime you plan to use.
  • This is an ablated research artifact, not a safety guarantee. Evaluate behavior before deployment.