Inference systems
Making models run faster on the machines people already own, from an eight-year-old GPU to a laptop, by measuring before we tune.
Flux-GPU · runtime v2Curiosity, into capability.
An independent AI research lab building more capable, useful, and reliable systems.
Our research
We study how AI systems behave on real hardware, for real people.
Every result we publish was measured, repeated and checked against a baseline. When something does not work, we say so.
Agents that finish the job
Assistants that plan, research, write code and use a browser, with every step visible and every send left to the person.
Astron · AstrixMeasurement and evaluation
Interleaved A/B runs, confidence intervals and hardware fingerprints, so a number means the same thing tomorrow as it does today.
ABBA protocol · Welch's t-testInteractive worlds
Games and experiences on Roblox, where players show us, every day, what people actually enjoy doing with technology.
NotFox gamesFlux-GPU: measured autotuning for local inference
A runtime that tries execution configurations on your machine, keeps only the ones that are statistically faster, and remembers them per hardware fingerprint. On an AMD RX 580 it made generation 68% faster.
Generation, tokens/sQwen3-4B Q4_K_M · RX 580 8 GB
+68.3%generation speed on an RX 580, 95% CI +67.9 to +68.8%
+39.7%generation on Qwen3-8B, found by the same search
12Blargest model run on an 8 GB card, 36 of 40 layers on the GPU
−7.3%time to first token on a 1,496-token prompt, through a live server
Latest
Latest papers and posts
Paper
Flux-GPU: measured autotuning for local inference on consumer GPUs
Screening, interleaved confirmation and per-machine profiles; results on an AMD RX 580 from 4B to 12B parameters.
Blog
One queue, 68% faster: what an autotuner found on an RX 580
The winning change was not a kernel or a quantization. It was which Vulkan queue the work was submitted to.
Blog
Measure, then tune: how we decide a change is real
Baselines first, interleaved runs, a confidence interval that has to clear 1%, and the word "inconclusive" when it does not.
The company