SenseNova — U1 MoT model and infographic artifacts¶

Development line: project:sensenova · thread u1-mot-model-and-infographic
Last event: 2026-08-01 · 1 dated since 2026-08-01 · Researched: 2026-09-06 · confidence: high

What it is¶

SenseNova U1 is an Apache-2.0 OpenSenseNova model family for practitioners running visual creation locally.

Abilities: - Text-to-image. - Image editing. - Interleaved image-text generation. - Visual understanding.

The current U1.5-8B-MoT card lists 18B BF16 parameters and no Hugging Face Inference Provider. Use U1.5 for general creation and editing, and Infographic-V3 only for infographic-specific work.

Development line¶

• 2026-08-01 — SenseNova linked U1 MoT preview and infographic artifacts on Hugging Face. On 2026-08-01, SenseNova added links to Hugging Face pages and Spaces for SenseNova-U1.5-8B-MoT-Preview and SenseNova-U1-8B-MoT-Infographic-V3. These links mark a model and demo milestone. They do not establish release status, technical changes, or artifact capabilities.

What changed¶

• 2026-07-16 — SenseNova-U1-8B-MoT-Infographic-V3 added integrated infographic generation and editing: local text/content plus global style/layout edits.
• 2026-08-01 — SenseNova-U1.5-8B-MoT-Preview was a July 31 Preview for broader generation and editing; it was paired with the earlier V3 card, not released together with it.
• 2026-08-04 — A community-maintained 19.9 GB Q8 GGUF made the Preview easier to run but did not change official weights.
• 2026-08-20 — SenseNova-U1.5-8B-MoT became the current official release; its matching 8-step LoRA added a faster path.
• 2026-09-01 — A community 21.2 GB Q8 GGUF appeared for the final U1.5 checkpoint.

How to use this¶

1. For a no-install evaluation, open the current U1.5 card's linked SenseNova-Studio playground; treat it as a workflow trial rather than production performance evidence. — https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT
2. For local use, clone OpenSenseNova/SenseNova-U1, run uv sync in its environment, and use the U1.5-8B-MoT model path with the text-to-image or image-editing entrypoint. — https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT
3. Choose a trained resolution bucket and configure device_map or the project VRAM mode before scaling resolution or batch size. — https://github.com/OpenSenseNova/SenseNova-U1/blob/main/examples/README.md
4. For infographic generation or editing, select SenseNova-U1-8B-MoT-Infographic-V3 rather than assuming the general U1.5 checkpoint has the same task specialization; the sensenova Space is an interactive V3 route. — https://huggingface.co/spaces/sensenova/sensenova-u1-infographic-v3

Best practices¶

• Use a direct prompt for a clear task with few constraints. Use Image PE for complex composition, copy, counts, layout, or reference planning, and vary seeds when diversity matters. — https://github.com/OpenSenseNova/SenseNova-U1/blob/main/docs/u1.5_best_practices.md
• For a complex edit, state the requested change and everything that must remain unchanged. Order multiple reference images deliberately and describe each image's role. — https://github.com/OpenSenseNova/SenseNova-U1/blob/main/docs/u1.5_best_practices.md
• Prefer the documented approximately 2K resolution buckets. Arbitrary width-height combinations can reduce quality. — https://github.com/OpenSenseNova/SenseNova-U1/blob/main/examples/README.md
• Inspect dense text, exact count/layout, anatomy, and broad multi-reference edits manually. Lower cfg_scale when detail or colour is over-emphasized. — https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT

Superseded by this¶

• 2026-08-20 — Treating SenseNova-U1.5-8B-MoT-Preview as the current general-purpose U1.5 checkpoint is obsolete.

Still unknown¶

• SenseNova is a broader SenseTime brand; the verified material identifies the OpenSenseNova U1 family but does not establish that U1.5 is the same product as commercial SenseNova 6.x or U1 Pro.
• The Preview demo is owned by hugging-apps rather than sensenova; its accessible page does not prove which checkpoint it currently runs, its uptime, or a production service commitment.
• The sources claim native 4K, but the checked examples document approximately 2K trained buckets and no independent 3840x2160 quality result for a specified production workload.
• No current hosted API pricing, SLA, or Hugging Face Inference Provider was verified for the final U1.5 checkpoint.

Sources¶

Agent brief¶

• Subject: project:sensenova, thread u1-mot-model-and-infographic, 1 dated events 2026-08-01 → 2026-08-01.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.