The GB10 Solution Atlas is now open source, the inference engine made for the community with breakneck inference speeds (Qwen3.6-35B-FP8 100+ tok/s)
Some of you saw our post a couple weeks back about hitting 102 tok/s stable on Qwen3.5-35B on a DGX Spark. A lot of you asked "cool, where's the code?" Today's the day: Github Atlas is open source. Pure Rust + CUDA, no PyTorch, …