Inference, reimagined for the agentic era
Frontier models, at breakthrough speed and efficiency, powered by an inference stack that improves itself with every model, chip, and request
Our research, including DFlash and AWQ, powers innovation at
Modal
vLLM
Decagon
- Ornith AI
The future of AI will be defined by the systems
that make models work at scale.
Introducing the inference stack
that improves itself
Our agents turn every new model, every new chip, and live production traffic into a faster stack, from GPU kernels to the serving engine, making frontier models faster, cheaper, and more reliable in production.
Autonomous agent loop
Every model, on every chip, refined by agents
Open-weight models, wherever you run them
Serverless per token, on your own compute, or on dedicated throughput for your team
- 01Inco PlatformBetaServerless open-weight modelsLearn more
- 02Inco SplashPreviewLocal AI on your own computeLearn more
- 03Inco EndpointsDedicated throughput for your teamLearn more
Every model, at the speed you need
- BatchHalf price, within 24 hours
- StandardThe list price
- FastFrom 2× the speed
Everything an agent needs from inference
Up to 99%+ cache hits
Agents resend the same growing context on every call. Repeats come back fast, at the cached price.
1M-token context
Agents carry long histories and whole codebases. Every model we serve keeps it all in view.
100% reference accuracy
Agents can’t afford a wrong answer mid-task. Faster never means different answers.
Your data, your rules
Agents handle your code and customer data. You decide how it is handled, all the way to zero data retention.
Explore the research powering our products
Our speed comes from research we publish in the open. DFlash models have been downloaded more than 10 million times on Hugging Face.
- #1 In output speed on Artificial Analysis, for every model we launched withInco AI launches its inference platform
- 3× Decoding speed with DFlash 2, at a third of the compute per tokenDFlash 2: Keep drafting parallel
- 4.4× Throughput of GLM 5.3 on launch day, vs. its native checkpointInco AI launches day-0 support for GLM 5.3
- 2× Splash vs. the next-fastest Mac engine we measuredSplash: A local engine built around the model
What people say
Inco is tackling the single biggest friction point in the expansion of agentic AI: the sheer cost and latency of running frontier models at scale. Lucas and Zhijian have translated world-class efficiency research into an autonomous production platform for inference.