Overview
SenseNova U1 (RiRiXin U1) is a series of "native understanding and generation unified models" officially released and **open-sourced** by SenseTime on April 28, 2026. Based on SenseTime's self-developed NEO-unify architecture from March this year, it is the first to **natively unify** multimodal understanding, reasoning, and generation capabilities within a single model architecture, moving away from the common "stitching" multimodal approach in the industry (i.e., simply stitching together understanding and generation models).
The initial open-source release includes two specifications of the Lite lightweight version: SenseNova-U1-8B-MoT (dense backbone) and SenseNova-U1-A3B-MoT (MoE mixture-of-experts backbone), along with weights, inference code, and technical reports. For developers, this is one of the few domestically open-sourced models in the first half of 2026 with original architectural breakthroughs.
**📌 2026 Supplementary Updates**:
- **2026/6/9 SenseNova Token Plan Public Beta Launch**: **All model calls are completely free during the period** (DeepSeek V4 Flash + SenseNova 6.7 Flash-Lite + SenseNova U1 Fast) — one of the few "top-tier model free" activities in China
- **2026/7/13 SenseNova-Vision Officially Released and Open-Sourced**: An important visual capability upgrade for the RiRiXin system, making vision a native capability of general foundation models, a key complement following U1
Key Features
- NEO-unify Native Unified Architecture: A single model unifies multimodal understanding, reasoning, and generation capabilities, rather than achieving them through multiple model stitching
- Fully Open Source: Weights, code, and technical reports are all open-sourced; enterprises can deploy privately with zero authorization barriers for secondary development
- Dual Backbone Options: The 8B dense version is suitable for low-VRAM deployment, while the A3B MoE version offers stronger capabilities at the same inference cost
- Unified Input/Output for Text, Images, Audio, and Video: Supports native multimodal interaction with images, text, audio, and video without the need for additional adapters
- Optimized for Chinese Scenarios: Specifically tuned for local scenarios such as Chinese understanding, OCR, document parsing, and image generation
- Dual-Channel Release on HuggingFace / GitHub: Weights on HuggingFace, code on GitHub, with ready-to-use inference examples and Docker images
Use Cases
- Enterprises and teams needing local/private deployment of multimodal large models
- Developers building multimodal retrieval, visual question answering, and text-to-image generation applications
- Universities and research institutions conducting multimodal architecture research
- Products requiring unified understanding and generation capabilities (e.g., both image understanding and generation)
- Teams seeking to move away from OpenAI/Anthropic commercial APIs for overseas or compliance-sensitive scenarios
Pros
- Original Architecture: NEO-unify is a rare underlying innovation in the domestic open-source ecosystem
- Fully Open Source: Complete set of weights, code, and reports, business-friendly
- Dual Specification Options: 8B suitable for single-GPU deployment, A3B MoE suitable for high-throughput services
- Native Unification: Image viewing, thinking, generation, and description within the same model, no stitching loss
- Friendly to Chinese and OCR scenarios, suitable for domestic business deployment
Pricing
The SenseNova U1 Lite series (SenseNova-U1-8B-MoT, SenseNova-U1-A3B-MoT) is open-sourced and free to use, with no licensing fees for commercial scenarios. SenseTime also provides a hosted API on the RiRiXin official platform, billed per token, with specific pricing subject to official announcements. For the MoE version, it is recommended to use mainstream inference frameworks (vLLM, SGLang, etc.) for optimal throughput.
Summary
SenseNova U1 is a heavyweight new member of the domestic open-source multimodal model landscape in spring 2026. Its biggest highlight is the NEO-unify native unified architecture—integrating understanding and generation into the same model rather than stitching them together. For teams needing multimodal capabilities but wanting to reduce reliance on overseas APIs, it offers a new option that is both open-source and architecturally original, especially suitable for Chinese scenarios and private deployment. For those seeking top-tier overall performance, Claude/GPT series can still be used as cloud supplements.