SenseNova-Loop-9B — 4bit MLX

An MLX conversion of sensenova/SenseNova-Loop-9B, supporting text and image inputs with text output on Apple silicon.

Precision and architecture

Eligible language projections and embeddings use 4-bit affine quantization, group size 64. The vision tower stays BF16. The mHC router weights and biases remain FP32 and match the source checkpoint exactly. Weight files occupy 5.963 GB (5.553 GiB).

The 32 physical decoder layers execute as 56 visits, each with an independent cache. The unused training-only loop gate is omitted.

Usage

Requires MLX-VLM with SenseNova Loop support (PR #2752). Install this validated implementation revision if your installed release does not include it:

python -m pip install -U "git+https://github.com/Neywa-Labs/mlx-vlm.git@3e90e9d9628ab386ff8dfbe561b09859cad9e3b9"
python -m mlx_vlm generate --model nativ-community/SenseNova-Loop-9B-4bit --prompt "What is 17 plus 25?" --max-tokens 128
python -m mlx_vlm generate --model nativ-community/SenseNova-Loop-9B-4bit --image image.jpg --prompt "Describe this image." --max-tokens 256
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("nativ-community/SenseNova-Loop-9B-4bit")
images = ["image.jpg"]
prompt = apply_chat_template(
    processor, model.config, "Describe this image.",
    num_images=len(images), enable_thinking=False,
)
result = generate(model, processor, prompt, image=images, max_tokens=256)
print(result.text)

Use enable_thinking=True for the model's thinking template. Multiple images use the ordered image list and num_images=len(images).

Conversion

Converted directly from the original floating-point checkpoint at revision 329a65f, using MLX 0.32.4:

python -m mlx_vlm convert --hf-path sensenova/SenseNova-Loop-9B --revision 329a65f701ba8c48b73c30d0cbaa6f0449733673 --mlx-path SenseNova-Loop-9B-4bit --dtype bfloat16 -q --q-bits 4 --q-group-size 64 --q-mode affine

Validation and limits

This export was reloaded and tested with greedy text generation and image description on an Apple M5 Max with 128 GB unified memory. The arithmetic check returned 42; the image check identified cats. All 96 mHC routing tensors remain byte-identical to the source FP32 tensors.

These are smoke checks, not a quality benchmark. The full 512K context limit has not been validated. The Qwen3.5 fused speculative verifier is unsupported and is rejected by the implementation.

Source

See the original model card for model details and upstream Apache-2.0 licensing information.

Downloads last month
-
Safetensors
Model size
9B params
Tensor type
U32
·
BF16
·
F32
·
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for nativ-community/SenseNova-Loop-9B-4bit

Quantized
(14)
this model