LLaDA2.2-flash-OptiQ-2bit

LLaDA2.2-flash-OptiQ-2bit

Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon (no PyTorch, no cloud). Try the Lab · All OptiQ quants · Docs

An OptiQ mixed-precision MLX quant of LLaDA2.2-flash, a ~100B diffusion language model with a 256-routed-expert sparse MoE. This is an extreme 2-bit build.

• Mixed 2/4-bit static build — per-layer bit-widths assigned to a 2.5 target bits-per-weight.
• 192 GB bf16 to 36 GB on disk (5.3x).

Requirements

Needs optiq >= 0.4.4, which ships the vendored llada2_moe decoder (the 256-expert diffusion MoE) and the block-diffusion decode loop. Stock mlx-lm has no llada2_moe arch and cannot load or generate from this repo.

pip install -U optiq

Running it

LLaDA2 is a masked-diffusion model, not autoregressive — it denoises a canvas block by block. optiq serve detects the arch and routes it through OptiQ's vendored decoder and the block-diffusion decode loop automatically:

optiq serve --model mlx-community/LLaDA2.2-flash-OptiQ-2bit

Then call the OpenAI-compatible endpoint at http://localhost:8000/v1, or use it from the OptiQ Lab and optiq code.