ZDTaichu5.0-9B MLX

MLX files for ZDTaichu5.0-9B, an advanced multimodal foundation model developed by the Zi Dong Tai Chu team for visual understanding, spatial reasoning, agentic tool use, and embodied AI workloads. Built on a Qwen3.5-9B language backbone and C-RADIOv4-H vision encoder, the source checkpoint supports a native context length of 131,072 tokens (128K).

Benchmarks

Benchmark results reported by TaichuAI for the original ZDTaichu5.0-9B foundation model. These figures are not measurements of this MLX conversion.

MLX files

Quantization	Repository	Size (GB)
8-bit	ZDTaichu5.0-9B-MLX-8bit	11.20 GB
6-bit	ZDTaichu5.0-9B-MLX-6bit	8.96 GB
4-bit	ZDTaichu5.0-9B-MLX-4bit	6.72 GB

Multimodal architecture

Component	Architecture	Precision
Language Backbone	Qwen3.5 hybrid (Gated DeltaNet + full attention, 3:1 ratio)	Quantized (8-bit affine, group_size=64)
Vision Encoder	C-RADIOv4-H (ViT-H/16, 653M)	BF16 (full precision)
Projector	RMSNorm → Linear(5120→20480) → SquaredReLU → Linear(20480→4096)	BF16 (full precision)

The vision encoder and multimodal projector are kept at full precision (BF16) to prevent visual reasoning degradation.

Chat template

The MLX conversion embeds the upstream chat template. An external copy is provided as chat_template.jinja for runtimes that require a separate template file.

Usage

Vision understanding with `mlx-vlm`

mlx_vlm.generate \
  --model abenzerps/ZDTaichu5.0-9B-MLX-8bit \
  --image path/to/image.jpg \
  --prompt "Describe what is shown in this image in detail." \
  --max-tokens 512 \
  --temp 0.7

Text generation with `mlx-vlm`

mlx_vlm.generate \
  --model abenzerps/ZDTaichu5.0-9B-MLX-8bit \
  --prompt "Explain why reproducible builds matter." \
  --max-tokens 512 \
  --temp 0.7

Python API

from mlx_vlm import load, generate

model_id = "abenzerps/ZDTaichu5.0-9B-MLX-8bit"
model, processor = load(model_id)

prompt = "Describe what is shown in this image in detail."
image = "path/to/image.jpg"

output = generate(model, processor, prompt=prompt, image=image, max_tokens=512)
print(output)

Increase context up to 131,072 tokens (128K) when sufficient unified memory is available on Apple Silicon. Tool-call behavior depends on the serving runtime and its parser integration; use the embedded template and verify tool calls in the target application.

Source and build

Source model: TaichuAI/ZDTaichu5.0-9B
Source revision: a22afd15a3f85659f103caa659ec4aa9500a998e
Conversion: ml-explore/mlx / ml-explore/mlx-vlm
Quantization format: MLX affine quantization (group_size=64, bits=8)
License: Apache-2.0

abenzerps/ZDTaichu5.0-9B-MLX-8bit