
Runs from yesterday aggregate decode hit 40.9/53.1/ 64.6 tok/s for C1-4. Warm TTFT measures 0.353 s at 8K and 0.361 s at 20K. KV pool scales to ~2.53M tokens FP4-local support is the best around too. Must try @jmacftw :)
✓Building Atlas, pure Rust and CUDA/ROCm kernels on DGX Spark & Strix Halo | MLPerf Agentic Edge Taskforce | Ambassador @Alibaba_Qwen | Ambassador @DevinAI

Runs from yesterday aggregate decode hit 40.9/53.1/ 64.6 tok/s for C1-4. Warm TTFT measures 0.353 s at 8K and 0.361 s at 20K. KV pool scales to ~2.53M tokens FP4-local support is the best around too. Must try @jmacftw :)

All of our SparkScore demos are real-time, recorded against a serving endpoint on one of our @NVIDIAAI DGX Sparks. SparkScore technology has been contributed to @AtlasInference - the O.G. of Rust-based inference engines. github.com/enntity/sparkscor…

Give early-access Qwen4 an @alibaba_cloud models crew like Image, Wan-2.7, Audio, Omni and see 🎥 It storyboarded, shot 10 clips, scored/re-scored after hearing a first mix, caught and fixed its own broken first cut, and signed off 🫡 @Alibaba_Qwen is cooking... can't wait 🍿

We started an open effort to train a DFlash2 drafter for GLM 5.3 Flash. It's already 92-95% as fast as one of the best drafters for the model and totally open licensed. Plug and play with most model engines and works great with our SparkGLM and @AtlasInference .

We have a 27B decision model reading a 12,000-token handbook on every request and making a decision in 191ms on one @NVIDIAAI DGX Spark. It follows the rules every time and the model hasn't been fine-tuned. Here's what we did. 🧵

“I wanna go fast!” - Ricky Bobby We got a very smart 27B decision model answering in 146 ms. On one DGX Spark. On your desk. The video shows real time serving of an actual endpoint. The possibilities are endless! @AtlasInference

SWE-2 is already fantastic, so excited to blaze through kernel dev with this flash variant 🔥

FP4-Local=better quality/BFCL accuracy, smaller footprint 🤏 TP=2, @Alibaba_Qwen's 3.8-Flash-Next 108 tok/s code, 260+ for 4 users 📈 Prefill ~2.3× vLLM, one 17K prompt, 5.8 vs 9.2s 💻 Worked w/ @jmacftw, >NVFP4 🧠 Try it: sparkrun run @atlas/qwen3.8-flash-next-fp4-local-tp2

Our community is so awesome ♥️ @Alibaba_Qwen Flash-Next orchestrated the whole thing with support from Voxtral for voices, Minimax for video and music, and Krea 2 for images 🎬 Running e2e through Atlas ☺️ youtube.com/watch?v=lKcHbDxj…

we brought @cognition to @GeorgiaTech 🚀 100+ students teamed up with @DevinAI for a 7 hour hackathon at @HackerHouseCo and shipped 25 projects. everyone left with devin credits and a pro plan for the month. - 1st place: AKA studio, describe a ring and get real CAD files back. akhilesh, armaan, greg, @Keylemx - 2nd place: Asymmetry, personal ai agents that connect your needs to your network - 3rd place: Raven, real-time emergency detection from traffic feeds and public alerts thank you @dabit3, azeez @atlasInference and the cognition team for working with us. :)


A break from the usual... as a @DevinAI ambassador, I co-hosted a hackathon at @GeorgiaTech, met the most cracked builders on campus 🔥 100+ students, 7 hours, and >25 shipped projects 🚀 Thank you @dabit3, @cognition, @HackerHouseCo 🐝

"I no longer think we're going to be able to just patch all the bugs." - Sam Altman We love @vllm_project, but 40 advisories in 6 weeks. 5 rated high, 3 let an attacker run code on YOUR server. Atlas is just Rust + CUDA/ROCm kernels, no dependencies. None of the 40 touch it :)

@RisingSayak would love to look into integrating this for GB10 as fused PiGemma (AdaRMS + gated residual/flow-matching step) so @LeRobotHF users get native kernelize() speedups out of the box!

Attention robotics fans 🚨 VLA's like @physical_int Pi05 are coming to Atlas 🤖 Watch Atlas control its first robot on GB10 solely through the model itself and the vision feeds ⚙️ DGX Sparks have many use cases beyond text! AGX Thors aren't the only physical edge devices 🫢

It's amazing to see +1 Rust-backed engine takeover :) Early 2026 Atlas came about as a pure CUDA zero dependency framework. I'd say we're worth a (re)try too. Just 30s to first token 😉 GLM5.3, Qwen3*, Nemotron3* github.com/Atlas-Inf/atlas sparkrun run @atlas/qwen3.8-27b-nvfp4