MiniMax

MiniMax has opened a public beta for M3.1-Flash, a lightweight native multimodal model in its M3 series. It supports a 1M-token context window and was trained from pretraining on mixed text, image, and video semantic spaces. In tests cited by the company, it generated teaching animations, procedural motion effects, interactive web pages, and pixel-style Canvas animations, with performance on several tasks matching Claude Opus 5.5; MiniMax also says it can proactively fill in design details.

Separately, MiniMax said its H3 video model has surpassed 24 million downloads in the month since being open-sourced, with more than 300 public derivative models.

An open-source acceleration stack for MiniMax's H3 video model replaces the default 50 denoising steps with 4–9 via a post-trained RH model. H3 Lightning, published on GitHub by RunningHub, pairs that with SageAttention2, Cache-DiT, torch.compile and TP2+Ulysses4 parallelism for PCIe multi-GPU hosts.

On four RTX 6000D GPUs, RunningHub reports a 5-second 1344×768 video drops from 348.8 seconds to 28.7 seconds, roughly a 12x speedup with BF16 precision retained; on eight GPUs a 15-second dual-reference video takes about 73 seconds. The GitHub release supports local multi-GPU deployment and does not state a licence.