Motif-3-Alis-MLX-4.5bpw

Motif-3 · Alis MLX 4.5bpw (speed tier)

Mixed-precision MLX build of Motif-Technologies/Motif-3 (314.8B total / 13.2B active MoE, MIT) for Apple silicon — the speed tier companion to avlp12/Motif-3-Alis-MLX-8bit. Decode is read-bandwidth-bound at batch 1, so halving the bits nearly halves the per-token read: ~50 tok/s with MTP self-speculative decoding (k=3) on a single M3 Ultra — about +35% over the 8-bit build at matching settings — at a small, measured quality cost (tables below).

Fits a 256 GB Mac with full headroom; a 192 GB M2/M3 Ultra fits with a raised wired limit and modest context.

Measured performance (M3 Ultra 512 GB)

Speed stack (all in the install fork): single-kernel Sinkhorn for the mHC gates, compiled activation/glue fusions (PolyNorm, router, mHC mixes, attention epilogue — bit-exact, kill switches MOTIF_COMPILE_ACT/GLUE=0), an mHC-transition mega-kernel (one threadgroup per position fuses the 16K rmsnorm, gate projections, Sinkhorn, premix, and input layernorm — ~7 serial dispatches to 1, +17%; MOTIF_MHC_TRANS=0), and vendor-faithful MTP self-speculative decoding (wiring confirmed by Motif in this discussion) with correct chained-draft norm handling. k=3 drafting wins on Apple silicon: batch-1 decode is latency-bound, so verifying 4 tokens costs only ~1.5x one token — the economics invert Motif's (compute-bound, vLLM) "1 speculative token is optimal" guidance. Leviathan rejection-sampling acceptance remains available for k=1 (MLX_MTP_REJECTION=0 to disable).

Quality vs the 8-bit build

Teacher-forced NLL and greedy top-1 agreement against the 8-bit reference on short KO/EN/code probes (same final-release weights; probe-scale — treat as indicative, not benchmark-grade. *top-1 measured on the shorter 300-token probe, on which the KO NLL delta read +9.9% — probe-scale variance; the 1.5k-token probe is the better estimate):

If you want maximum fidelity (e.g. as a distillation teacher), use the 8-bit build. Motif themselves ship a 4-bit NVFP4 checkpoint for vLLM, and report the MTP head keeps ≥70% draft acceptance at 4-bit — consistent with what this build measures.

Quantization recipe

C6 mixed profile: experts 4-bit g64, attention/dense/shared-expert 6-bit, wkv_a 8-bit (latent chokepoint), router gate / mHC / lambda_proj / norms / PolyNorm coefficients in bf16. MTP head 8-bit (model-mtp.safetensors, +254 MB). Per-tensor map in config.json. Tensors >2³¹ elements sanitized without mx.split (mlx#3836 workaround, byte-verified).

Install & run

pip install git+https://github.com/avlp12/mlx-lm.git@motif3-support
mlx_lm.generate --model avlp12/Motif-3-Alis-MLX-4.5bpw \
  --prompt "한국의 전통 건축양식을 요약해줘." --max-tokens 400 --temp 0.7 --mtp --mtp-num-draft-tokens 3
# OpenAI-compatible server:
mlx_lm.server --model avlp12/Motif-3-Alis-MLX-4.5bpw --port 8003 --mtp --mtp-num-draft-tokens 3

Usage notes

• Sampling: temp ≈ 0.6–0.8. At temp 0 the think channel can enter repetition loops.
• --mtp-num-draft-tokens 3 is optimal on Apple silicon (latency-bound verify makes wide drafting cheap; k=2 is within noise, k=1 is ~−20%). Motif's own "1 speculative token" guidance applies to compute-bound GPU serving, not this regime.
• Greedy transcripts under MTP can diverge from the plain path after many tokens (ULP-class tie flips between batched-verify and single-step kernels); the sampling distribution is equivalent.