Mage-VL-8bit

microsoft/Mage-VL (4B codec-native multimodal model, Apache-2.0) quantized to 8-bit for MLX on Apple Silicon.

Converted with mlx_vlm.convert (affine, group size 64, 9.023 bits/weight): the Qwen3-4B language model — embeddings, attention/MLP projections, lm_head — is int8; the Mage-ViT vision tower stays bf16. 5.0 GB on disk vs 9.5 GB upstream.

Quality

Gated against the bf16 checkpoint and the PyTorch reference on the model's own examples/dog.jpg, greedy decoding:

comparison	result
int8 vs PyTorch reference	48/48 tokens identical
int8 vs bf16 MLX	48/48 tokens identical
final-position logits vs reference	cos 0.9982 (bf16 baseline: 0.9996)

Measured on an M5 Max (macOS 27), one process per number, floor read post-load, at the default 4096-token budget:

	resident floor	activation	decode
int8 · image	5.0 GB	3.0 GB	85.7 tok/s
int8 · video 16f	5.0 GB	5.8 GB	80.0 tok/s
bf16 · image	8.85 GB	2.85 GB	51.2 tok/s
bf16 · video 16f	8.85 GB	5.8 GB	50.2 tok/s

Activation is unchanged by quantization (5.77 vs 5.79 GB) — only the resident floor moves.

The token budget, not the frame count, is the memory lever: at a fixed budget, 32 frames costs slightly less than 16 because per-frame resolution drops to compensate. Raising the budget is what costs — the same clip at an 8192-token budget needs 14.2 GB of activation.

Corrected 2026-07-29. This section previously read "~14 GB for 16-frame video QA at a 4096-token budget". That 14 GB figure is real but belongs to the 8192-token budget: it was measured before per-frame video resolution was derived from the token budget, and was not re-measured when the default changed. Swift package v0.3.0 re-declares its footprint accordingly.

Use

Python — the mage_vl architecture is merged into mlx-vlm (PR #1745, merged 2026-07-29), so no fork or branch install is needed:

pip install "mlx-vlm>=0.6.9"
# 0.6.9 is the first release containing mage_vl. Until it lands on PyPI, install from main:
#   pip install git+https://github.com/Blaizzy/mlx-vlm

No trust_remote_code needed — the port ships a torch-free processor, and no torch or torchvision install is required.

If you installed from the old xocialize/mlx-vlm@mage-vl branch, switch to upstream — that branch is retired and will not receive fixes. Anything installed from it before bf9ca87 (2026-07-28) also carries a bug where torchvision-free setups silently fell back to the checkpoint's torch-based remote processor and failed with ones_like(): argument 'input' must be Tensor. Details in discussion #1.

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

model, processor = load("mlx-community/Mage-VL-8bit")
prompt = apply_chat_template(processor, model.config, "Describe this image in detail.", num_images=1)
print(generate(model, processor, prompt, image=["dog.jpg"], max_tokens=64).text)

Swift — xocialize/mage-vl-swift (≥ v0.3.0) serves this checkpoint as an MLXEngine package (imageAnalysis + videoAnalysis):

try await engine.register(
    MageVLPackage.registration,
    configuration: MageVLConfiguration(quant: .int8))

Provenance

Upstream weights: microsoft/Mage-VL, Apache-2.0. This repo excludes the DCVC neural-codec runtime and the StreamMind gate weights (CUDA-only / not used by the MLX ports). The neural_codec Python sources are retained verbatim from upstream for trust_remote_code completeness.