Modelpublic

mlx-community/Tmax-9B-MLX-4bit

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

1likes64downloads

Model Card

Tmax-9B MLX (4bit)

MLX-converted text-only weights of `allenai/tmax-9b`.

The upstream base ships as a multimodal Qwen3_5ForConditionalGeneration config but contains zero vision tensors in its safetensors — i.e. it is already a text-only checkpoint with stub vision metadata. This release strips the residual vision_config / image-token entries so it loads cleanly via mlx_lm without a vision tower.

—Source: allenai/tmax-9b
—License: Apache-2.0
—Variant: 4bit
—Quantized by: raullenchai
—Tooling: mlx-lm 0.31.3 (the upstream mlx_vlm 0.3.12 qwen35 loader hard-requires vision-tower weights that this base does not ship, so the text-only `mlxlm.convert` path is used instead)
—Chat template: ships with the source repo (chat_template.jinja)
—Tool format: qwen3_xml-compatible (<tool_call>{json}</tool_call>)

Usage

python

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Tmax-9B-MLX-4bit")
print(generate(model, tokenizer, prompt="Hello", max_tokens=32))

Notes

—This is a pure text-generation MLX release. No vision/image inputs.
—For best chat behavior, use the chat template that ships with this repo.

Benchmarks

Measured on M3 Ultra Studio (28 (20 Performance and 8 Efficiency) CPU, 60-core GPU, 256 GB unified memory) via rapid-mlx 0.8.18. Medians of 3 runs.

Variant	Decode tok/s	TTFT (ms)	Prefill 1k (tok/s)	Prefill 4k (tok/s)	Prefill 16k (tok/s)	Tool-call e2e
Tmax-9B (4-bit MLX)	107.4	127	1,060	1,124	1,092	726 ms (OK)

Recommended default for the 9B family on M3 Ultra — ~19% faster decode than the Qwen3.5-9B-4bit control on the same hardware (90.5 tok/s), tool-call e2e under 1 s.

Full results (all 7 Tmax MLX variants + 2 Qwen3.5 controls): rapid-mlx docs.

Reproduce:

bash

pip install rapid-mlx==0.8.18
rapid-mlx serve tmax-9b --port 8765