SenseNova-U1.5-8B-MoT-8step-v2-4bit

Apple Silicon (MLX) artifact of SenseNova-U1.5-8B-MoT — SenseTime's unified T2I + editing + VQA flagship (NEO-unify Mixture-of-Transformers, pixel-space rectified flow, no VAE).

This artifact is the official 8-step distillation LoRA, V2 (SenseNova-U1.5-8B-MoT-LoRA-8step-V2, released 2026-09-24, mirrored at xocialize/SenseNova-U1.5-8B-MoT-LoRAs), α/rank-merged into the generation stream, plus 4-bit (group 64) stream quantization. Quantization: group size 64, 4-bit (mlx convention quantization block in config.json).

V2 vs V1. Upstream describes V2 as its recommended 8-step adapter: better colour balance, more natural contrast and less oversharpening than the V1 adapter in SenseNova-U1.5-8B-MoT-8step-4bit. Measured here at fixed seed over five prompts (1024², 8 steps, cfg 1): mean saturation −0.071, RMS contrast −0.049, 99th-pct gradient −30 %, shadow clipping 8.5 pp lower, highlight clipping 14.8 pp lower; composition is preserved (cos(V1, V2) 0.80–0.96). int4 sits at cos 0.976 / 24.2 dB against the V2 bf16 merge-at-load render (tighter than the V1 tier's ~0.92). The V1 artifact stays published unchanged so existing seeds keep reproducing.

Runtime: sensenova-u1-swift (MLX-Swift), variant 8step-v2-4bit. Weights are stored in the runtime's key layout (NHWC convs; quantized linears as weight/scales/biases) and load with zero conversion transient — peak load memory ≈ resident.

Performance (M5 Max)

peak 14.7 GB · 1024² in 3.2 s

Use it for: the smallest, fastest tier for T2I and editing — fits 16 GB-class Macs. Note: 4-bit weight quantization on a diffusion path measurably perturbs the denoise trajectory: at a fixed seed it produces a different but equally valid image. Prefer the 8-bit tier when bf16-faithful output matters.

Editing and VQA run on this tier too (runtime ≥ 0.5.1, 2026-10-11): the distill adapter touches only the 294 generation-stream projections, so VQA uses the base tier's understanding stream unchanged (answers match token for token), and the 8-step cfg-free edit matches the 50-step base edit (a 2048² jacket edit in 15 s vs 169 s, 28.7 dB apart). Think mode stays a base-tier feature. Evidence: Docs/receipts/8step-v2-ab/ and AB-R-0490 in the runtime repo.

Quick start

git clone https://github.com/xocialize/sensenova-u1-swift && cd sensenova-u1-swift && swift build -c release
hf download mlx-community/SenseNova-U1.5-8B-MoT-8step-v2-4bit --local-dir SenseNova-U1.5-8B-MoT-8step-v2-4bit
.build/release/sensenova-cli --weights SenseNova-U1.5-8B-MoT-8step-v2-4bit \
  --prompt "A cinematic mountain lake at sunrise, realistic photography." \
  --width 1024 --height 1024 --steps 8 --cfg 1.0 --out out.npy

The runtime also does instruction editing (--edit-image) and VQA (--vqa) on this tier, and <think> reasoning mode (--think) on the base tiers — see the repository README for the full surface and the parity report (component parity < 1e-4; e2e per-pass cosine 0.999+ vs the reference PyTorch implementation).

Provenance & license

• Upstream: sensenova/SenseNova-U1.5-8B-MoT @ 07d76f6, Apache-2.0, by SenseTime / SenseNova — paper · reference implementation.
• Adapter: sensenova/SenseNova-U1.5-8B-MoT-LoRAs @ f33b8fe, file SenseNova-U1.5-8B-MoT-LoRA-8step-V2.safetensors (sha256 f4190cabd01fa261a45be853275c483d3a8069295c52cf01f7b2b31f4356f201), Apache-2.0. Merged in fp32 as W += (α/r)·BA over the 294 gen-stream projections and cast; against the V1 artifact exactly 882 tensors differ (294 × weight/scales/biases) and nothing else.
• This repository is a format conversion (official-LoRA merge + bf16 cast + 4-bit group-64 quantization) of the upstream checkpoint, redistributed under the same Apache-2.0 license with modifications noted here. All credit for the model to the SenseNova team.
• tokenizer.json is generated from the upstream vocab.json/merges.txt (byte-identical tokenization, verified against reference ids).