Atlas is open source!
An inference engine written from scratch in Rust + CUDA. No PyTorch, no Python, no 200-dependency install dance. <2 min cold start, BUILT for GB10
Qwen3.6-35B at 130 tok/s on a single DGX Spark.
Demo powered by @Gradio
🧵 https://t.co/vxZLwBJMub
@loktar00@pupposandro You can get 30-40tk (or even more) with the new dflash speculative decoding techniques. I’ve been using 27b on one of my Sparks at 40-60 tok/s with DFlash.
@loktar00@pupposandro This is an unfair comparison because no one really uses ollama as a backend on the Spark. the standard is vLLM for its concurrency/parallelism support. It’s way faster and more robust.
Hi @cerebras + @cline , this is my entry for the hackathon.
Enter GLM Token Saver
https://t.co/YqN51ce6vs
Uses Speculative Decoding (tested on DGX Spark) to save on precious Cerebras Inference Tokens!
@redtachyon@RayFernando1337 There is an unsloth version and a fix to llama.cpp. There’s also an update coming up that fixes flash attention, it’s running at 65tk/s initially on my spark