Tess-4-27B is a thinking-native, agentic 27B model built on Qwen/Qwen3.6-27B, post-trained on 64K-token long-context agentic traces plus a reasoning style distilled from a three-model teacher ensemble (Opus-4.8, GPT-5.5, GLM-5.2). It's multimodal (inherits Qwen3.6's vision tower — text and image input) and reasons via explicit <think>...</think> blocks before answering.

Architecture


Base model	Tess-4-27B (Qwen3.5-VL / `Qwen3_5ForConditionalGeneration` arch, ~27B params, vision + video capable)
Quantization	MX FP8 (micro-scaled 8-bit float)
Bits per weight	8.381
Disk size	27 GB (from ~52 GB bf16)
Format	MLX (safetensors)
Minimum unified memory	Needs 32GB+ unified memory. Quantization completed cleanly (clean file output, correct bpw), but on the 24GB Apple M4 Pro used for this conversion, loading the model for generation hit a Metal `kIOGPUCommandBufferCallbackErrorOutOfMemory` — the 27GB of weights plus KV cache/activations exceed what a 24GB machine can allocate to the GPU. Not smoke-tested locally; expected to run fine on higher-memory Apple Silicon (32GB/48GB/64GB+).

Usage

python

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "sahilchachra/tess-4-27b-mxfp8-mlx"
model, processor = load(model_path, trust_remote_code=True)
config = load_config(model_path, trust_remote_code=True)

prompt = apply_chat_template(processor, config, "Explain the difference between TCP and UDP.", num_images=0)
out = generate(model, processor, prompt, image=None, max_tokens=1024, verbose=False)
print(out.text if hasattr(out, "text") else out)

For image input, pass an image path/URL and set num_images=1 in apply_chat_template, and pass the image to generate(..., image=<path_or_url>).

CLI:

bash

mlx_vlm.generate --model sahilchachra/tess-4-27b-mxfp8-mlx \
  --prompt "Explain the difference between TCP and UDP." \
  --max-tokens 1024

This model uses the Qwen3.5-family chat template with explicit <think>...</think> reasoning blocks — allow a generous max_tokens budget (1024+) so generation doesn't cut off mid-thought.

Smoke test

Not run locally — see the RAM note above. Quantization itself completed without error (correct file layout, correct bits-per-weight reported by mlx_vlm convert). If you run this on a 32GB+ Mac and hit issues, please open a discussion on this repo.

Other MLX variants

Credits

Base model: migtissera/Tess-4-27B by Migel Tissera
Quantization tooling: mlx-vlm