This is
nex-agi/Nex-N2-mini converted
to MLX format and quantized with oMLX's oQ4 mixed-precision scheme (4-bit base,
sensitivity-driven bit allocation, MSE-optimal clipping). The result is about 19 GB
on disk, roughly 4.7 bits per weight effective, with sensitive tensors such as the
linear-attention projections kept at higher precision.
Nex-N2-mini is an agentic model built around what its authors call Agentic Thinking:
it interleaves reasoning, tool use, and environment feedback rather than treating them
as separate stages. The architecture is a hybrid MoE (qwen3_5_moe): 40 layers
alternating linear attention with full attention every fourth layer, 256 experts with
8 active per token, and a 262k-token context window.
The original checkpoint includes a vision tower. MLX text inference does not use it,
so the vision weights were dropped during conversion; this copy is text-only. Expect
around 21 GB of memory in use during inference.
It also works out of the box with oMLX.
Tool calling works without any extra configuration. The chat template uses the
Qwen3-Coder XML style, which mlx-lm and oMLX both detect automatically, so servers
return proper structured
tool_calls, and thinking ends up in the reasoning field
instead of leaking into the response content. Tested end to end with
Swival as the harness, including multi-step tasks that
exercise file edits, search, and shell commands while the model is thinking.
An 8-bit companion is available at
jedisct1/Nex-N2-mini-mlx-8bit.
