Qwen3.8-Flash-Next Is on TensorFold with Speed Boosts (x.com)

TensorFold has unveiled its latest inference engine, significantly enhancing the performance of Qwen3.8-Flash-Next on DGX Spark systems. The engine boasts impressive decode speeds of over 62 tokens per second for single streams and 119 for five concurrent streams, alongside an astounding prefill rate of 2,500 tokens per second. With a robust 256k context window and improved performance in time-to-first-token and multi-user scenarios, this upgrade marks a substantial leap over previous vLLM configurations. Additionally, its compatibility with both vision and video inputs on a unified memory architecture opens new avenues for multimodal applications. This advancement is particularly noteworthy for the AI/ML community as it democratizes access to high-performance machine learning capabilities, allowing effective use on consumer-grade hardware like the RTX 3090 equipped with 64GB RAM. The community is excited about potential optimizations for other models, such as GLM 5.3 Flash, which may further enhance the breadth of applications and accessibility of AI technologies. This progress underscores the ongoing trend of making sophisticated AI tools more available to developers and researchers, fostering innovation in various fields.