🔥 You don’t need to re-read a long doc for every follow-up anymore. TensorFold… · @QwenAIFans · TwKit

@QwenAIFans

🔥 You don’t need to re-read a long doc for every follow-up anymore. TensorFold 1.1.5 just dropped — and the Qwen numbers are wild. Same 35k-token document, new question: • Flash Next on a GB10: first token 25.7s → 2.8s • Qwen3.8-27B on an M3 Ultra: 125.9s → 9.0s It keeps the prompt state instead of prefilling the whole thing again. Flash Next also got faster cold: up to +28% prefill on GB10, and +16–23% decode on Macs. Caveat: TensorFold’s own benches on their NVFP4 / MLX builds — your stack may differ. via GitHub release 👇 https://github.com/ashhart/TensorFold/releases/tag/v1.1.5

曝光 413 · 评论 2 · 点赞 6 · 书签 1