Google's open TimesFM-3 forecasts your data with zero training
The 330M-param model tops forecasting benchmarks and runs on a laptop — but ships non-commercial; meanwhile vLLM cracks fast inference on AMD GPUs.
Copy markdown
Zero-shot forecasting, no training loop
TimesFM-3 is a 330M-parameter foundation model for multivariate time series: point it at your history and get forecasts with no fine-tuning. Google reports it tops zero-shot forecasting benchmarks, and the weights are up on Hugging Face — but under a non-commercial license, so check terms before it reaches production.
Fast open-model inference without NVIDIA
vLLM's speculative decoding now runs on AMD Instinct MI300X and MI355X. DFlash and MTP draft methods reach up to 2.87x on Gemma 4 and 2.68x on Kimi-K2.5 — pushing Gemma 4-26B from 2,344 to 6,434 tokens/sec. n-gram and Medusa aren't wired up yet, and gains shrink on smaller models like Qwen3-8B.