Nunchux introduces VC-Attention, a training-free low-bit attention method that accelerates video generation by 1.6× on NVIDIA B200 compared to BF16 FlashAttention-4. The technique combines V-Smooth for reducing quantization error and ExpCast-FP8 for eliminating the softmax bottleneck, achieving higher fidelity than existing methods like SageAttention2.
MiniMax H3 Max Prompts is a curated collection of 30 real-world examples demonstrating text-to-video prompting techniques across 7 creative categories, with copyable prompts, breakdowns, and step-by-step guidance for users to adapt and reproduce the results.
Independent research on safety mechanisms in large-scale Mixture of Experts language models across 9 models and 200+ experiments reveals that safety functions as a holographic attractor state rather than deletable circuits, with three independent pathways (attention, routing, residual) that must be neutralized simultaneously. Key findings show safety decisions commit early in processing, quantization affects steering differently based on method, and structural modifications are fundamentally impossible in 300B+ reasoning models.