MiniMax | Hailuo 3: First open-source mainstream multimodal model [CSC AI Internet]

English summary

1. Open-source is the core of this release: Hailuo 3 becomes the first open-source model among mainstream multimodal models, with capabilities comparable to leading closed-source models. According to ArtificialAnalysis blind test arena, H3 text-to-video (including audio) ranks 2nd globally with Elo 1242, and video editing (including audio) ranks 1st. H3 brings open-source video models into the first tier for the first time. 2. Pricing continues to offer high cost-performance: H3 generates 2K video at about 0.13 per second, which is 1/3 to 1/2 of peers. We believe Hailuo 3 can further lower the penetration rate of multimodal usage. 3. The previous generation Hailuo 2.3 was released in October 2025. This time, after 10 months, H3 is a cross-generation upgrade: resolution increased from 1080p to native 2K/24fps, single generation extended from max 10 seconds to max 15 seconds, new 'one-shot output' feature, can feed up to 9 reference images + 3 videos + 3 audio clips simultaneously, and supports instruction-based local editing.