Onyx-maple-preview-2bit

MLX 2-bit repack of deepgrove/maple-preview (20B-total / ~1B-active ternary-weight MoE reasoner, MIT license) for the Onyx app's own inference engine (OnyxLLM).

The repack is lossless for every projection. The source BF16 master weights are per-row ternary {-s, 0, +s}; naive min/max 2-bit affine quantization cannot represent the zeros, so each row is packed manually as MLX affine 2-bit with scale = s, bias = -s (dequantization reproduces the BF16 master bit-exactly; verified per tensor during conversion). Attention q/k/v/o and all 3×256 expert projections per layer are packed this way (168 tensors, all lossless). Embeddings and lm_head are continuous in the source and are quantized 8-bit affine (group 64). The fp32 router gates and norms stay BF16.

Experts are fused per layer into stacked mlp.switch_mlp.{gate,up,down}_proj tensors ([256, out, in]) for single-gather_qmm MoE dispatch.

Conversion: Onyx session 2026-08-05. Not affiliated with deepgrove; see LICENSE (MIT) for the upstream terms.