sensenova-u1-8b-infographic-v3-mlx

SenseNova-U1 8B Infographic V3 — MLX quant-matrix (SceneWorks re-host)

Pre-quantized MLX turnkey tiers of sensenova/SenseNova-U1-8B-MoT-Infographic-V3, hosted for the SceneWorks app. V3 is a tensor-identical checkpoint refresh of the NEO-Unify architecture (config + all 1,116 tensor keys byte-identical to V2), independently trained with stronger image editing.

Tiers (each subdir is a self-contained load root)

• q4/ (default) — packed Q4 backbone (~11.8 GB)
• q8/ — packed Q8 backbone (~19.9 GB)
• bf16/ — dense source mirror (~35 GB)

The backbone decoder-stack Linears (attention + SwiGLU, both understanding & generation paths) are packed into one model.safetensors; the token embedding, lm_head, norms, vision convs, and the flow-matching head stay dense. The loader packed-detects ({base}.scales) — no in-app re-quantize. The bf16 tier is the dense source mirrored.

Provenance

• Source: sensenova/SenseNova-U1-8B-MoT-Infographic-V3 (Apache-2.0)
• Converter: mlx-gen-sensenova::convert::prequantize_turnkey (SceneWorks/inference)
• Default inference: 50 steps, CFG 4.0, timestep-shift 3.0 (image edit img_cfg 1.0)

License: Apache-2.0 (inherits the source license).