SenseTime tagged SenseNova-U1.5-8B-MoT as an official release on GitHub on August 20, under an Apache 2.0 license, with weights synced to Hugging Face and ModelScope. That's three weeks after the July 31 preview version, and less than four months since the product line's first open-source release — U1-8B-MoT on April 27 this year.
The new version leads with two things: native 4K image generation, and folding image editing and precise control into the same model.
One model, three parameter counts
Start with the part that's easy to get confused by. The model's name says 8B. The repo README adds a note that "8B-MoT" refers to roughly 8 billion understanding parameters plus roughly 8 billion generation parameters. The Hugging Face model card, meanwhile, lists the tensor scale as 18 billion, in F32 and BF16 precision.
All three numbers are correct — they're just counting different things. MoT stands for Mixture of Transformers, the architecture SenseTime uses to keep understanding and generation on separate towers that alternate on the same sequence. Whether to count the two towers together, report only the activated scale, or list them separately has no industry convention yet; every company's release page for its unified multimodal model handles it differently.
The practical effect for developers: comparing these models by parameter count alone is becoming less useful. A real comparison needs to know whether the model fits on a single GPU, how long it takes to produce a 4K image, and peak VRAM usage — none of which SenseTime gave a unified figure for on the release page.
Direct 4K generation, not upscaling after the fact
The "native 4K" claim is worth unpacking. There are currently two common paths to a 4K image: one has the model generate a lower-resolution image first, then hand it to an upscaler to regenerate the detail; the other has the generation model work directly at the target resolution.
The difference between the two isn't the final pixel count — it's structural consistency. With the upscaling route, the global composition is locked in at the low-resolution stage, and the upscaling step can only add detail. That makes it prone to misalignment when the image has dense text, a complex layout, or relationships between multiple subjects. SenseTime built 4K into the generation step itself this time; the company's own wording is that the model achieves "overall composition and extremely fine texture and micro-detail at once," and it stresses that 4K generation efficiency has improved in this version.
The release page lists six user-visible improvements. Aside from 4K, the other five all revolve around controllability: composition, materials, and lighting closer to real photography; native image editing that better preserves subject identity and spatial structure; improved readability for Chinese-English text rendering and complex layouts like infographics and posters; steadier adherence to complex instructions involving object count or spatial relationships; and the ability to use detection boxes, visual markers, and multi-image references for precise control.
That last one is the most useful for teams building design tools. Pointing at a region with a box or marker and saying "change this" is faster than repeatedly rewriting a prompt to describe a position, and it's easier to wire into an existing editor interface.
Putting the known flaws on the release page
The back half of the model card has a section listing five unresolved issues: details tend to get over-enhanced, dense text still produces errors, performance is unstable under constrained layouts, character details are inconsistent, and complex edits can drift.
It's not common for release materials from a domestic Chinese open-source model to volunteer a section like this. Its value isn't just that it reads as honest — these five points are, roughly, the shared open problems for this class of model right now. Writing them down saves downstream users a round of trial and error, and gives future versions a baseline to be measured against.
SenseTime also released SenseNova-U1.5-8B-MoT-LoRA-8step alongside the main model, a 0.4B, eight-step distilled version aimed at latency-sensitive use cases. Combined with the Apache 2.0 license — commercial use allowed, no copyleft clauses — and hosting across GitHub, Hugging Face, and ModelScope, this release lowers the barrier to access about as far as it can go.
What's left comes back to the same old constraint: open weights solve whether you can get the model, not whether you can run it. An 8B-plus-8B dual-tower structure generating 4K images directly isn't going to be light on VRAM, and how many teams actually put it to use will still come down to whether a single GPU is enough.
Sources: CocoLoop, GitHub repository OpenSenseNova/SenseNova-U1, Hugging Face model card, IT Home; version tag, license, parameter figures, and the six improvements were cross-checked against the repo README and model card.