DeepSeek
Collection
DeepSeek V4 Flash AXQ 2/4/6-bit MTP, Flash-0731 2/4-bit MTP plus MXFP4/6-bit stubs, Pro-0813 hobby 2-bit MTP, and DeepSeek-OCR-2. 3-bit withdrawn. • 10 items • Updated
How to use AutomatosX/AX-DeepSeek-V4-Flash-0731-MLX-AXQ-MXFP4 with MLX:
# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm
# if on a CUDA device, also pip install mlx[cuda]
# Generate text with mlx-lm
from mlx_lm import load, generate
model, tokenizer = load("AutomatosX/AX-DeepSeek-V4-Flash-0731-MLX-AXQ-MXFP4")
prompt = "Once upon a time in"
text = generate(model, tokenizer, prompt=prompt, verbose=True)How to use AutomatosX/AX-DeepSeek-V4-Flash-0731-MLX-AXQ-MXFP4 with MLX LM:
# Install MLX LM uv tool install mlx-lm # Generate some text mlx_lm.generate --model "AutomatosX/AX-DeepSeek-V4-Flash-0731-MLX-AXQ-MXFP4" --prompt "Once upon a time"
AXQuant MXFP4 pack of DeepSeek-V4-Flash-0731 (--q-mode mxfp4).
Recipe examples/deepseek-v4-experimental-mxfp4-v0.1.yaml.
Checkpoint Tier 1 is not certified on df-macstudio-m2 (192 GB). If the
pack is in the 170 GB+ class, factory generate/cert is memory-blocked.
This is an MLX Apple MXFP4 pack, not an OCP or NVIDIA checkpoint.
Base model
deepseek-ai/DeepSeek-V4-Flash-0731