3.9×
Sooner to the target
An agent tuning FlashAttention-3 reached its target at iteration 14 with our trace, against 54 with the best existing tracer.
From silicon to API to applications
MarsCompute serves open models through one API, builds the tools that trace and tune GPU kernels, and ships AI-native products on top: one stack for the industries, enterprises and tasks that run on AI.
Products
From the GPU kernel to the finished application, each layer is a product you can use on its own.
Open text and video models through one OpenAI-compatible API.
Kernel tracing as a service: upload a CUDA kernel, get back a timeline of what happened inside it.
The toolbox for agentic GPU kernel optimization: tracing, binary analysis and microbenchmarking.
A multi-stage AI review that makes a research paper reviewer-ready.
Technology
Our optimization work is a loop, not a one-off tuning pass. Each turn leaves a faster kernel and a record of what worked, which the next agent reads before it starts. That is recursive self-improvement, applied to AI infrastructure.
A timeline of every warp inside the compiled kernel, recorded without changing the kernel.
A coding agent reads the trace and finds the stage that is waiting, and what it is waiting for.
The agent changes the kernel: a pipeline stage, a schedule, a parameter.
The new kernel is measured and checked for accuracy. What works is kept, and the loop starts again.
3.9×
An agent tuning FlashAttention-3 reached its target at iteration 14 with our trace, against 54 with the best existing tracer.
+13.3%
Over the open-source kernel, after an agent studied traces of a closed-source one. 5.2 to 13.3% across shapes.
+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
19
NVIDIA from Volta to Blackwell, and AMD across CDNA and RDNA, traced the same way.
Token Factory
Already using an OpenAI SDK? Change the base URL and the key, and keep the rest of your code.
Chat completions and streaming in the format your SDK already
speaks. To switch models, change the model
field.
A single key reaches every text and video model. Limit a key to one model when you want to, and revoke it at any time.
Prepaid credit from $10, charged per request at the published price. A request that fails before answering is not charged.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.marscompute.ai/v1",
api_key=os.environ["MARSCOMPUTE_API_KEY"],
)
reply = client.chat.completions.create(
model="qwen3-30b-a3b-instruct-2507",
messages=[
{"role": "user", "content": "Count from one to five."},
],
max_tokens=256,
)
print(reply.choices[0].message.content)
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.marscompute.ai/v1",
apiKey: process.env.MARSCOMPUTE_API_KEY,
});
const reply = await client.chat.completions.create({
model: "qwen3-30b-a3b-instruct-2507",
messages: [
{ role: "user", content: "Count from one to five." },
],
max_tokens: 256,
});
console.log(reply.choices[0].message.content);
curl https://api.marscompute.ai/v1/chat/completions \
-H "Authorization: Bearer $MARSCOMPUTE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3-30b-a3b-instruct-2507",
"messages": [
{"role": "user", "content": "Count from one to five."}
],
"max_tokens": 256
}'
Text models take a chat request. Video models return an MP4.
A flagship open model for coding, knowledge work and long tasks.
k3
LiveZ.ai's flagship open model, built above all for coding.
glm-5.3
LiveDeepSeek's large V4 model for reasoning, coding and agents.
deepseek-v4-pro
LiveA reasoning mixture-of-experts model for coding and agents.
minimax-m3
LiveText, or a first and last frame, to a 2K clip of 5 to 15 seconds, with optional sound.
minimax-h3
LiveA prompt and up to 12 image, video and audio references to a clip of 3 to 15 seconds with sound. Powered by MiniMax H3.
pika-h3-flex
Seventeen more text models, among them GLM-5.3 Flash, DeepSeek V4.1 Flash, Qwen3.8-27B and Kimi K2.6, are on the Models page.
Solutions
Whether you call a model or tune a fleet, the same stack is behind it.
Hyperscalers, AI labs and hardware makers. At fleet scale, a few percent of kernel throughput adds up to thousands of GPUs.
More from every GPUOpen models behind one API, with spend that never runs past your credit and usage you can audit.
AI across the businessReview a manuscript before a reviewer does, and run experiments on open models without a contract.
ResearchMeta
Under a commercial collaboration, MarsCompute provided Meta with a customized G-Watch to help its engineers optimize internal AI workloads.
Read the storyDevelopers
Guides for the API, and real traces to learn the tools from.
Sign up for the API in minutes, or talk to us about your own kernels, hardware and workloads.