I started INT21 two months ago, and I’m proud to announce that we’re coming out of stealth today with our first product: PTX Kernel Factory. We’ve created the first self-improving agent swarm for producing expert-level PTX kernels autonomously and at massive scale. https://int21.ai

INT21 | Self-Improving Compute Infrastructure

Source: https://int21.ai/

Use compute to improve compute.

INT21 builds self-improving AI systems for the software beneath modern AI. Our first product generates and optimizes low-level GPU software, then proves its work with tests and benchmarks.

INT21Use compute to improve computeINT21Use compute to improve compute PTX Kernel FactoryFirst public releaseJune 16, 2026Company launch

Introducing INT21 and PTX Kernel Factory

We are building self-improving AI systems for the software beneath modern AI. The first four implementations produced by PTX Kernel Factory are open source, and the product is entering beta.

01 / First proof

Generated by PTX Factory. Measured against best baseline.

PTX Kernel Factory produced four GPU kernel artifacts across Hopper and Blackwell. We tested them on matching hardware and inputs against established expert implementations, with correctness verified before timing.

GH200 / Hopper1.59x

Peak measured performance.

B200 / Blackwell1.52x

Optimized integration.

GH200 / Hopper8.17%

Faster on geometric mean.

B200 / Blackwell126 / 126

Faster in every comparable case.

Operator-level benchmark results, not full-model speedup claims.

Autonomous optimization loopEvidence retained across generationsPTX Kernel Factory autonomous improvement loopAn autonomous agent swarm generates GPU kernels, evaluates them on target hardware, retains the evidence, and feeds that knowledge into the next generation.PLANWRITEREVIEWTUNEAGENTSWARMAUTONOMOUSTARGET GPULIVE01COMPILE02CORRECTNESS03BENCHMARKBEST VALID CANDIDATE+18.7%GENERATIONSGEN 01GEN 08GEN N+1SEARCH MEMORYRETAINED EVIDENCE STARTS THE NEXT GENERATIONPTX Kernel Factory autonomous improvement loopAn agent swarm generates kernels, target hardware evaluates them, and retained evidence improves the next generation.AUTONOMOUS SWARMPLANWRITEREVIEWTUNEAGENTSWARMTARGET GPULIVECOMPILECORRECTBENCHMARKBEST VALID CANDIDATE+18.7%RETAINED EVIDENCEGEN 01GEN 08GEN N+1EVIDENCE COMPOUNDS INTO THE NEXT GENERATION

1. 01### Fully autonomous swarm Specialized agents plan, implement, review, and optimize each kernel end to end.
2. 02### Grounded in real hardware Every candidate is compiled, verified, and benchmarked on the target GPU.
3. 03### Improvement compounds Reusable evidence improves both the search process and the kernels it produces.

Cloud-native swarm control planeLive system model1. 01 / Elastic scaleThousands of agentsCloud-native scheduling expands the swarm around available compute. 2. 02 / Shared directionOne measurable goalEvery agent works against the same constraints and acceptance criteria. 3. 03 / Generational memoryExperience carries forwardResults, failures, and strategies become the next generation’s starting point.

1. 01Agents
2. 02Models
3. 03GPU
4. 04Infrastructure
5. 05Cloud

AI-native operating model

Engineering capacity scales with compute, not headcount.

Our experts set direction, constraints, and acceptance criteria. Autonomous agent swarms execute, evaluate, and retain the work, so adding compute expands how much engineering INT21 can perform.

Founders / Cross-stack operators

Research, systems, and infrastructure experience carried into one company.

### Bing Xu

Founder & CEO

• Agents
• Models
• GPU

Bing co-authored the original Generative Adversarial Nets paper, created XGBoost’s Python package, and co-created MXNet and AITemplate. Before founding INT21, he was a Distinguished Engineer at NVIDIA following its acquisition of HippoML, the GPU inference company he founded.

### Qingye Jiang

Founding Partner

• Infrastructure
• Cloud

Qingye has spent more than a decade building and tuning high-performance computing and distributed systems at AWS. His work spans workload analysis, performance engineering, cloud infrastructure, and real-time systems.

PTX Kernel Factory / Beta

Bring us a 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