Atlas is an open source LLM engine we hand tuned for NVIDIA DGX Spark. One ~75 MB binary, no Python, no PyTorch. What ships is what we verify, and we bench it every single release.
curl -fsSL https://atlasinference.dev/quickstart.sh | sh Do not take our word for it. First token in under 90 seconds on a DGX Spark. Median of our GB10 runs, model cached, atlas 59616dc, Jul 2026. Same command below, run it and time it yourself.
Three things landed this month and every card links straight to the primary source.
MLCommons published the new MLPerf Inference v6.1 edge agentic benchmark and Atlas Inference is named as a contributor alongside NVIDIA. It measures multi turn agentic LLMs on a single edge accelerator, BFCL v4 for accuracy and replayed agentic coding trajectories for performance. We helped shape it because it is the benchmark that actually looks like the work.
Read the MLCommons announcement →Big thanks to AMD for the Strix Halo desktop. We took Atlas to ROCm to show what this silicon can really do when it is paired with custom kernels, and that desktop is the box we ran and submitted MLPerf on.
See the post on X →Atlas is submitted to MLPerf Inference v6.1 in the closed edge division, the same CUDA source across NVIDIA GB10 and AMD gfx1151. llama.cpp is the bar we measure ourselves against on this workload and we like where we landed. Results stay under embargo until MLCommons publishes, so stay tuned.
Follow along in Discord →Atlas went from one Reddit post to a whole crew of builders running it on their own Sparks. The curve below is live, regenerated from the GitHub API on every deploy.
“Night and day compared to the 10 minute torch.compile cycle. Startup in about 15 seconds and it just stays coherent in an agentic loop.”
“Testing Atlas on a DGX Spark in an agentic workflow for over an hour. Super impressed. Spark is actually awesome with Atlas.”
“I had grown tired of the usual stack and was hoping for something like this. Really surprised and impressed. So glad I bought a Spark.”
Hundreds of builders are running Atlas on their own Sparks right now. We are in there every single day, shipping fixes, taking model requests, and tuning kernels live. Your machine is the test fleet and your voice sets the roadmap. Pull up.
The website is a build artifact of the repo. Models come from recipes, performance comes from committed gate enforced baselines, stamped with commit and date. If a number is not in the repo, it is not on this page.
An Atlas image ships only after the serve matrix passes: every model boots, stays coherent (greedy determinism, no token leakage, tool reliability), and holds throughput within 10% of its committed baseline. What “verified” means · gate_results.py
A release that ships slower than the committed baseline fails our gate. That one sentence is the whole positioning.
Submitted to MLPerf Inference v6.1 in the closed edge division, the same CUDA source across NVIDIA GB10 and AMD gfx1151. Results are under embargo until MLCommons publishes them, so stay tuned.
Atlas is a member of MLCommons and sits on the Edge LLM taskforce, where we helped shape the new v6.1 edge agentic benchmark. MLCommons names Atlas Inference as a contributor in the announcement. read the announcement.
The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See mlcommons.org for more information.
python3 tests/run_all_models.py && python3 tests/gate_results.py --update-baselines Beat these numbers or catch a regression, open an issue and we will feature it.
Every model card comes from a recipe in atlas-recipes.
Our MLPerf Inference v6.1 submission is in, closed edge division, on both GB10 and gfx1151. The numbers render right here in this receipt the moment MLCommons publishes them, gate enforced, reproducible, stamped. Until then the release gate holds every image to liveness and coherence, and you can reproduce any run yourself.
One multi model binary serves a full matrix of hand tuned targets on a single GB10. NVFP4 and FP8, MTP speculative decoding, EP=2 across two Sparks. Every target passes the serve matrix before we cut an image.
One codebase, both camps. Our CUDA kernels compile straight for AMD gfx1151 with SCALE by Spectral Compute. No HIP port, no second kernel tree. AMD sent us a Strix Halo desktop and that is the box we ran and submitted our MLPerf Inference v6.1 numbers on.
Pick a vendor, then a family. Every card maps to one recipe in atlas-recipes, so the site cannot list a model we do not ship. Copy the command and run it as is. Qwen3.6 leads because it is our flagship.
sparkrun run @atlas/qwen3.6-27b-fp8 --hosts localhost sparkrun run @atlas/qwen3.6-27b-fp8-mtp --hosts localhost sparkrun run @atlas/qwen3.6-27b-nvfp4 --hosts localhost sparkrun run @atlas/qwen3.6-27b-nvfp4-prefill-record --hosts localhost sparkrun run @atlas/qwen3.6-27b-nvfp4-unsloth --hosts localhost sparkrun run @atlas/qwen3.6-35b-a3b-fp8-bf16head --hosts localhost sparkrun run @atlas/qwen3.6-35b-a3b-fp8-mtp --hosts localhost sparkrun run @atlas/qwen3.6-35b-a3b-fp8-nvfp4head --hosts localhost sparkrun run @atlas/qwen3.6-35b-a3b-nvfp4 --hosts localhost This is the first 60 seconds. Everything after, per model recipes, EP=2, tuning, lives in the docs.
$ curl -fsSL https://atlasinference.dev/quickstart.sh | sh # checks for sparkrun, installs via uvx if missing, runs the flagship recipe
The script checks for sparkrun, installs it with uvx if missing, then runs the flagship Qwen3.6 recipe.
Rather not pipe curl to a shell. Install sparkrun, then run the flagship recipe direct.
uvx sparkrun setup install sparkrun run @atlas/qwen3.6-35b-a3b-fp8-mtp --hosts localhost The first 60 seconds live here. Everything after, per model recipes, EP=2, tuning, lives in the docs. Read the deployment guide · README
AI worth having should run on hardware you own. Prosumer machines like DGX Spark and Strix Halo are the first generation that makes that real, and we build for them first.
Pure Rust because the whole stack should be inspectable by one person, HTTP to kernel dispatch, no interpreter in the hot path. We develop on machines granted by NVIDIA and AMD. Both camps handed us silicon and we intend to continue proving to the world the raw power of these machines.
Open to all. The test fleet is the community desks. If a model matters to you, it matters to us.
Atlas grows from the desks it runs on. Every path below is real and linked. Contributions ship in the Community Edition under AGPLv3, and the CLA lets us re license for the Enterprise Edition.
Boot the matrix on your own GB10 and report what you see. Regressions and wins both get featured.
Deployment guide →Recipes are the model SSOT. Add a model, tune a quant, open a PR against atlas-recipes.
atlas-recipes →Hand tuned attention, MoE, GDN, Mamba-2 for Blackwell. Register level work, no generic fallbacks.
Good first issues →Improve the guide, triage issues, or just tell us what you are running in Discord.
Discussions →Contributions are AGPLv3 and the CLA permits Enterprise re licensing. See CONTRIBUTING.md.
Everything real links to an issue, a PR, or the Discord where the work happens. The teasers are teasers, and we say so.
Three GB10s in one rig for the really big models. More memory, more experts, more headroom. We are wiring up the topology now.
Talk trifecta in Discord →Active conversations with Intel about bringing Atlas to the Arc Pro B70. The email chain is live and we are waiting on confirmation. Nothing signed yet, but we are fired up about it.
Follow along in Discord →Native gfx1151 through SCALE. AMD sent us a Strix Halo desktop and we took Atlas to ROCm on it, custom kernels and all.
PR #187 →The same CUDA source submitted across GB10 and gfx1151, closed edge division. No numbers until MLCommons publishes.
Read the benchmark announcement →Our fused Gated DeltaNet kernel for Qwen3.6 landed in Hugging Face Transformers.
transformers #46423 →Large MoE NVFP4 ports across EP topologies, DeepSeek and Kimi class, tracked in the open.
Open issues →Building on Spark or Strix, bringing hardware to the table, or wanting to partner or talk business. We want to hear from you and we move fast.
Running Atlas in production or eyeing the Enterprise Edition. Tell us what you need and we will get you sorted.
Frameworks, benchmarks, standards bodies. If it makes local AI better we are all in.
Got silicon you want Atlas running on. Send it our way and watch what we do with it.