NVIDIA GH200
584 upstream tests and 235 package tests passed.
Product PTX Kernel Factory
Private beta / NVIDIA GPUs
PTX Kernel Factory
GPU kernels that improve with every run.
A fully autonomous agent swarm explores implementations, tests them on target NVIDIA hardware, and carries what it learns into the next generation.
- 01 Explore Parallel agent swarm
- 02 Build Compile candidates
- 03 Prove Target GPU evidence
- 04 Remember Advance the generation
Generation memory Useful evidence never returns to zero.
01 Compounding system
The factory does not reset.
Each run expands the evidence available to the swarm. Over time, the search becomes more informed, the strategies become more effective, and the resulting kernels become stronger.
- 01
Explore in parallel
Specialized agents search different implementation paths at the same time.
- 02
Measure what survives
Candidates are compiled, checked, and benchmarked on the target hardware.
- 03
Carry the evidence forward
Useful results become experience for the next search, not a discarded trace.
02 Inside SwarmOS
See the factory think.
SwarmOS exposes every level of a run, from the global control plane to one agent's reasoning, plan, and workspace.
03 Open-source proof
Published work, not a promise.
The first four factory artifacts cover RMSNorm and Kimi Delta Attention across Hopper and Blackwell. Source, tests, benchmark methods, and reports are open for inspection.
NVIDIA B200
All 580 upstream tests passed.
NVIDIA GH200
48 package tests and 65 subtests passed.
NVIDIA B200
Faster in every comparable case; all 66 package tests passed.
Operator-level results do not imply the same full-model speedup. The launch article includes test coverage, limitations, and measurement context.
04 Beta access
Choose how your team enters the factory.
Starter and Pro include subscription access plus discounted spot GPU access. Tokens and cloud compute are billed as used.
Starter
$200 / month
- 25% off spot GPU access
- Pay as you use for tokens and cloud compute
- Data may be used to improve our product
Pro
Recommended plan$2,000 / month
- 50% off spot GPU access
- Pay as you use for tokens and cloud compute
- Data will not be used
Enterprise
Contact us
- Talk with the INT21 team about enterprise access.
Bring us one hard GPU workload.
Start with an operation that is too slow, a new architecture without a mature kernel, or an important workload that has not justified weeks of specialist time.
Request beta access