jedisct1/Nex-N2-mini-mlx-OptiQ-8bit-MTP

Nex-N2-mini, OptiQ 8-bit MLX, native MTP

This is nex-agi/Nex-N2-mini converted to MLX format, quantized with oMLX's oQ8 mixed-precision scheme, and extended with a multi-token prediction head for native MTP speculative decoding in oMLX (enable the Native MTP toggle in the model settings).

About the MTP head: the released Nex-N2-mini checkpoint declares an MTP layer in its config but does not ship the weights. Since Nex-N2-mini is post-trained from Qwen3.5-35B-A3B-Base and keeps identical dimensions, this conversion grafts the MTP head from Qwen3.5-35B-A3B onto the Nex-N2-mini trunk. MTP drafts are always verified by the main model, so outputs are exactly what the trunk would produce on its own; the head only affects speed. In practice the grafted head reaches around 40-50% draft acceptance on prose and reasoning, less on unusual token sequences, for up to roughly 1.5 tokens per backbone pass.

Note that the preserved mtp tensors confuse the weight-name heuristics of plain mlx-lm, so this repo is intended for oMLX. For mlx-lm, use jedisct1/Nex-N2-mini-mlx-8bit instead, an 8-bit quantization without the head.

The original vision tower is not included; this copy is text-only.

Tool calling works without any extra configuration, including through the MTP decode path. The chat template uses the Qwen3-Coder XML style, which oMLX detects automatically, so the server returns 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.

A 4-bit companion with the same MTP head is available at jedisct1/Nex-N2-mini-mlx-OptiQ-4bit-MTP.