Inference acceleration
Generate the whole response in parallel, not one token at a time
Inception replaces sequential autoregressive generation with diffusion-style parallel decoding: every output position is initialized at once, then refined jointly across K iterations, instead of one token per forward pass.
How parallel decoding works
Parallel seed
All output positions are initialized simultaneously in a single forward pass. The model begins with a complete draft of every token at once, rather than producing them one by one left to right.
Joint refinement
Each of the K refinement iterations updates every position jointly, so tokens can attend to and influence each other across the full output length throughout the decoding process.
Collective convergence
The model converges when all positions reach sufficient confidence together. The output is not a left-to-right stream but a jointly decoded sequence refined until the whole is coherent.
Built for real workloads
Where parallel decoding fits
Interactive dialogue
Conversational assistants
Chat applications where response completeness matters more than streaming individual tokens. Inception produces the entire answer jointly, so the delivered output is globally consistent rather than left-to-right committed.
Structured generation
Documents and code
Long-form outputs such as reports, summaries, and code files that require coherent cross-sentence and cross-function dependencies. Parallel refinement lets late positions inform early ones before anything is committed.
Scale-out pipelines
High-volume batch jobs
Workflows generating large numbers of independent completions simultaneously. The compute constraint shifts from sequential token dependencies to batched parallel decoding, enabling denser utilization of available hardware.
Limited early access
Join the inference research program
We are opening API access to a small group of applied AI teams whose workloads are long enough and latency-sensitive enough to see the difference. Tell us what you are building.
Request AccessThe team
The people building Inception
Stefano Ermon
CEO & Co-Founder
Stefano co-founded Inception in 2023 to apply deep generative modeling research to the practical problem of inference latency. His focus is on the decoding side of language model serving.
Yuna Park
Head of Research, Co-Founder
Yuna's research background is in diffusion models for discrete sequences. She works on the refinement schedule and token posterior estimation that make parallel decoding tractable at production scale.
Marcus Dubois
VP of Engineering
Marcus leads the systems side of Inception: GPU kernel work, distributed serving, and the batching layer that keeps parallel decoding throughput-efficient across request shapes.