3.9×
Sooner to the target
An agent reached its target at iteration 14 with our trace, against 54 with the best existing tracer.
Why MarsCompute
Most companies in AI own one layer: a model, an API, or a tool. We build the layers in between, and that is why each of them gets better.
Four reasons
Applications, an OpenAI-compatible API, an optimization system and the tools that observe the GPU, built by one team. A problem found at one layer can be fixed at the layer that causes it.
See the stackAgents trace, diagnose, rewrite and verify GPU kernels in a loop. Each turn leaves faster infrastructure for the next: recursive self-improvement, one layer below the model.
How the loop worksEvery performance number on this site comes from a report we have published, and is listed with the hardware, the shapes and the cases where it did not help.
Read the resultsPrepaid credit from $10 and a published price per token. Spend never runs past your credit, and a request that fails before answering is not charged.
See pricesThe stack
The top two layers are what you call. The layers below are what we bring to the models we run on our own GPUs.
Layer 01
Paper Review Studio is the first: a multi-stage review of a research paper, run on Token Factory models. Building on our own stack is how we find what the layers below need.
Layer 02
Twenty-one text models and two video models are live. One key reaches all of them, cost is reserved against your credit before a request is sent, and every request is metered.
Layer 03
It discovers optimization opportunities from how workloads really execute, models how kernels, memory and runtime interact, and plans changes that do not conflict with each other.
Layer 04
G-Watch combines intra-kernel tracing, binary analysis and microbenchmarking, for engineers and the coding agents working beside them. MarsTrace offers the tracing as a service, on a GPU we bring.
Layer 05
NVIDIA from Volta to Blackwell, and AMD across CDNA and RDNA. Kernels written in CUDA, HIP, CuTe DSL, Triton, TileLang or FlyDSL, including closed-source vendor libraries.
Proof
3.9×
An agent reached its target at iteration 14 with our trace, against 54 with the best existing tracer.
+13.3%
Over the open-source kernel, from traces of a closed-source one. 5.2 to 13.3% across shapes.
94–98%
Original instructions still in a kernel traced by Xtrace. Existing tracers keep 8 to 48%.
+50.8%
Best case from an automated run; 25 to 51% at sequence length 1,024 and up with batch size 4 and up.
Report, March 2026
Meta
Under a commercial collaboration, MarsCompute provided Meta with a customized G-Watch to help its engineers optimize internal AI workloads.
Read the storySign up for the API, or talk to the optimization team about your own kernels and hardware.