SenseNova U1.5

What it is

SenseNova U1.5 is an open-weight multimodal model from SenseTime. A preview version focused on native 4K image generation appeared on 31 July 2026, and the main release, SenseNova-U1.5-8B-MoT, followed on 20 August 2026^[1]. It is an 8-billion-parameter mixture-of-transformers model that handles understanding, reasoning and generation in one network, without external vision encoders or VAEs^[2]. The weights are released under the Apache License 2.0^[1].

In plain terms, many multimodal systems are assembled from specialist parts: a vision encoder that reads an image, a language model that reasons about the result, and a separate decoder that draws the answer. U1.5 keeps all of that inside a single set of parameters^[2].

How images are handled

Instead of passing an image through a separate encoder, U1.5 tokenizes it directly: each 32x32 pixel region of the picture becomes one visual token^[2]. Under that scheme, a 4096x4096 image is a 128x128 grid of regions, so the number of tokens grows with image area — which is why the tokenization scheme matters more as resolution rises.

For generation, U1.5 uses a new spatial decoder in place of the per-patch MLP head used in U1, and this is what supports native generation up to 4096x4096^[2]. Native here means the model produces the image at that resolution itself, rather than generating something smaller and enlarging it afterwards.

Reported results

On the text side, SenseTime reported 86.67 on MMLU-Pro, 90.41 on C-Eval, 93.35 on IFEval and 69.00 on IFBench^[2]. MMLU-Pro and C-Eval test knowledge and reasoning questions; IFEval and IFBench test instruction following — whether a model completes a task in the form and under the constraints asked for.

On image generation it reported GenEval 0.92, DPG-Bench 88.11 and CVTG-2K 0.948^[2]. SenseTime described the GenEval score as the best among open-source models, said the DPG-Bench score ranks among the best-performing open-source models, and said the CVTG-2K score was the best overall among all the models it evaluated^[2]. GenEval looks at how well a generated image matches the text prompt, so it is one indicator of how well a model draws what was asked for. All of these are the developer's own comparisons; the evaluation setup, the models compared against and the conditions for reproduction belong to the technical report.

Report and follow-ups

The technical report was published on 11 September 2026. SenseTime said it would open-source the training code for supervised fine-tuning (SFT), reinforcement learning (RL) and on-policy distillation^[1][2]. Together with a permissive licence, that means users get not only the weights but a route to the pipeline that makes them usable^[1]. A LoRA-8step-V2 variant of SenseNova-U1.5-8B-MoT was released on 24 September 2026 with improved visual quality^[1].

Where it sits

U1.5 is the successor to SenseNova U1^[1], and the direction is visible in the design: seeing, reasoning and drawing handled by one set of parameters rather than by glued-together specialist modules, with weights and training code opened so that outside researchers can inspect, fine-tune and compare^[2].