MiniMax Open-Sources H3 Video Model—China’s AI Video Play
MiniMax open-sources H3, its first video model—builders get a new lever against ByteDance and Google. Plus: new research tools.
The story of the day
MiniMax’s H3 is China’s latest gambit to disrupt the AI video market. By open-sourcing its flagship model, the firm hands developers a free alternative to ByteDance and Google’s proprietary offerings—no strings attached. The move pressures Western incumbents to either match the openness or double down on performance. For builders, it’s a rare win: a high-capability video model without vendor lock-in or usage fees. Expect rapid forks, fine-tunes, and the usual scramble to verify claims of parity with closed rivals.
New tools & launches
H3 — Open-source text-to-video model from MiniMax, targeting ByteDance/Google competitors. [The Information]
Models & research
Recursive Transformers for Semiconductor Reliability — Proposes hardware-aware transformer architectures to replace expensive simulations in chip design, trading parameter bloat for compute efficiency. Open-weight. [arXiv:2607.27251]
DoTime — Synthetic benchmark generator for interventional/counterfactual time-series causal inference, with PyPI package and frozen evaluation suites. Open-source. [arXiv:2607.27263]
RLPF for Code Generation — Reinforcement learning framework to train code agents to prefer faster correct implementations, addressing runtime as a fragile reward. [arXiv:2607.27271]
Industry pulse
OpenAI’s latest blog post frames its European compliance efforts as a proactive alignment with the EU AI Act—useful for builders tracking regulatory risk, though light on technical detail. [OpenAI Blog]
Stat of the day
0 — Number of prior open-source video models from MiniMax before H3. The firm’s pivot to openness marks a strategic shift in China’s AI video race. [The Information]
Get the full analysis — subscribe to Gradient Pro.
