Anthropic released FlashProve, an Apache 2.0 zkSNARK prototype enabling cryptographic verification that an AI response came from a specific committed model without exposing model weights (MarkTechPost). The system demonstrated inference at 110 tokens per second on a single H100 GPU with 19 MB proofs, addressing verification and authenticity challenges in production deployments.
Tencent’s WeChat Vision team released WeVisOne-VL, a 2B and 4B open-weight vision language model under Apache 2.0 licensing that unifies detection, optical character recognition, segmentation, and alpha matting in a single architecture with a 9.6M-parameter pixel decoder (MarkTechPost). Separately, NVIDIA demonstrated three stacked optimizations for mixture-of-experts inference on DGX Spark, achieving 42.5 percent throughput improvement on Ant Group’s 124B Ling-3.0-flash model with no detectable accuracy loss (MarkTechPost).