A 400,000 CNY prize pool is ready! Let’s solve omni-modal large model deployment on Huawei Ascend’s domestic compute ecosystem
The AI industry is seeing a concentrated burst of intelligence capabilities. Human-computer interaction is moving from single-modal text conversation to real-time multimodal collaboration, and omni-modal models have become the focus of this leap. They break down the information barriers between text, images, audio, and video, achieving deep cross-modal fusion and understanding within a unified architecture, giving AI true native multimodal intelligence. For this reason, their demands on inference efficiency, compute adaptation, and engineering stability far exceed those of pure text models.
ModelBest's MiniCPM-o 4.5 is a leader in this direction. This omni-modal foundation model for real-time interaction supports end-to-end unified modeling of vision, speech, and text, delivering a natural "see, listen, and speak proactively" interaction experience, while continuously exploring low-latency, natural, and sustainable omni-modal application experiences. For omni-modal models to truly enter application scenarios, strong model capability is not enough — efficient, stable, and reproducible inference deployment and implementation solutions are also required.
That is exactly the challenge the MiniCPM & Ascend Inference Optimization and Application Innovation Challenge wants to tackle with developers. Initiated by ModelBest OpenBMB together with Huawei Ascend and MindSpore, the challenge centers on the MiniCPM-o 4.5 omni-modal model and invites university students, AI developers, and AI application innovation teams to explore efficient inference, low-latency interaction, and real-world application on Ascend NPUs, producing technical results that are runnable, verifiable, and demonstrable.
The competition focuses not only on whether the model "runs," but on first-response speed, end-to-end latency, concurrency, resource utilization, engineering reproducibility, and application experience in real interaction scenarios. Developers are encouraged to turn model capabilities into experienceable App/Web demos covering a range of deployment directions. Whether you work deep in AI Infra and domestic compute adaptation, or want to build a genuinely useful, demonstrable AI application, this competition is worth joining — with a 400,000 CNY prize pool waiting for you!
Two parallel tracks, one of them is yours
The competition accepts individuals or teams of no more than 3 people, with two tracks: high-performance inference optimization and innovative applications. All entries must be reproduced and verified in a unified Ascend environment to ensure fairness and reproducibility. The former focuses on the efficiency, stability, and reproduction quality of inference services; the latter, based on runnable demos, evaluates productization of omni-modal capabilities, interaction experience, and engineering completeness.
Track 1: High-performance inference optimization
If you are familiar with AI Infra, inference deployment, compiler optimization, or kernel optimization, and are interested in the MindSpore and Ascend developer ecosystem, this track is for you. Participants must complete inference adaptation and efficiency optimization for MiniCPM-o 4.5 in an Ascend NPU environment.
Scoring focuses on:
Model inference adaptation
TTFT / first-response speed
Single-chunk latency
E2E latency
Throughput and concurrent sessions
Resource utilization and stability
Accuracy loss control
Deployment and reproduction quality
Final submission:
Code/config, benchmark scripts, performance report, and reproduction instructions.
Track 2: Innovative applications
The innovative applications track is aimed at AI product developers, multimodal interaction designers, student innovation teams, and application engineering teams. Participants must use MiniCPM-o 4.5's omni-modal capabilities to build their ideas into a runnable, demonstrable, and experienceable demo.
Encouraged directions:
Real-time Q&A
Companion assistant
Visual/voice interaction
On-device applications
Multimodal scene understanding
Application scenarios in education, office, creation, and life services
Final submission:
Runnable demo / Web demo / App, PPT, project description, and demo video.
Awards: inside the 400,000 CNY prize pool
The total prize pool is 406,000 CNY (pre-tax). Awards are set separately for the two tracks, with 12 winners in total. If the number of eligible winners is insufficient, the corresponding awards and prize amounts will be reduced based on the actual review results.
High-performance inference optimization track
Innovative applications track
Additional notes: 1. All prize amounts are pre-tax; actual disbursement is subject to applicable tax regulations. 2. Winning entries must meet the competition rules, submission requirements, and reproduction requirements. 3. The final list of winners is subject to the review results published by the organizers.
Key dates
Competition notes
The competition is open to individual developers and teams; teams may have no more than 3 members.
A live leaderboard is available; all entries must pass reproduction verification in the unified MindSpore/Ascend environment to enter the valid ranking.
Each team may submit at most 3 entries per day.
For more detailed competition rules, follow the official website.
If you want to push the boundaries of large-model inference efficiency, or build a truly experienceable, deployable, and demonstrable omni-modal AI application, then you are the strongest builder this competition is looking for. Click the link below to join this omni-modal AI inference and application innovation challenge!
➤ Official registration page
🔗 ascend.openbmb.cn