Volodymyr Kuleshov's diffusion LLM guide maps the road from masks to models

Why it matters

Kuleshov is turning years of Cornell research into a development stack for Inception's commercial bet: that parallel text generation can make AI agents faster and cheaper without surrendering too much quality.

Volodymyr Kuleshov (@volokuleshov), the Cornell professor and Inception co-founder, published a technical guide on July 5 that lays out how researchers can turn masked-token prediction into a working diffusion language model. The document puts a coherent development path around a field that has spent the past two years moving from academic papers into products from Inception, Google and NVIDIA.

Kuleshov wrote the guide with Cornell researchers Marianne Arriola, Yair Schiff and Guanghan Wang. It was adapted from lectures and workshop talks delivered earlier in 2026, making the article a consolidation of existing work rather than a new model release.

That distinction matters because the guide is also a category-building document for Inception. Kuleshov and his co-founders are asking developers to reconsider the left-to-right generation method behind mainstream language models. Explaining the alternative in enough detail to reproduce it helps turn diffusion from a product claim into an engineering choice.

Kuleshov has followed this research-to-product path before. He earned his computer science Ph.D. at Stanford, where his thesis received the Arthur Samuel Best Thesis Award. He later co-founded Afresh, which applies machine learning to grocery inventory and food waste. According to his Cornell profile, Afresh is deployed in about 10% of U.S. supermarkets. Earlier genome-sequencing work was commercialized through Stanford spin-off Moleculo and became part of Illumina's genome-phasing service.

From a blank page to a model that revises itself

The guide begins with the central constraint of an autoregressive LLM: it generates one token after another and cannot revisit a token once it has been emitted. A mistake near the beginning of a response becomes part of the context used to generate everything that follows.

A diffusion language model starts with an entire sequence of masks or noisy tokens. It predicts many positions together, keeps some predictions, corrupts or masks others, and repeats the process. Later passes can use context from both sides of a token and revise earlier choices.

The simplest training recipe resembles BERT. A model receives a sequence with a random share of tokens masked and learns to reconstruct the original. Varying the masking rate across training examples turns that familiar objective into a generative process: generation begins with every position hidden and gradually fills the sequence.

That basic system still lacks several properties needed in production. It assumes a fixed output length, can require too many denoising passes, and cannot automatically repair every committed token. Much of the guide concerns the engineering work required to remove those constraints.

Block diffusion handles variable-length output by generating groups of tokens sequentially while refining the tokens inside each group in parallel. Remasking allows the model to hide a previous prediction and try again. Uniform-state diffusion replaces tokens with other vocabulary entries rather than a mask, giving the model a natural way to edit any position. Distillation trains a model to compress multiple refinement rounds into fewer steps.

Schiff and Wang have worked on guidance, remasking and inference-time correction methods. The guide packages those separate research threads into something closer to a technical stack.

Inception is selling the latency argument

Inception's commercial pitch follows directly from that stack. Its Mercury models generate multiple tokens per pass, shifting more of the decoding workload into the parallel computation that GPUs handle well. Inception markets Mercury 2 for reasoning and Mercury Edit 2 for latency-sensitive coding workflows.

Inception says Mercury delivers a time to first token below 300 milliseconds, throughput five to seven times higher than speed-optimized frontier models, and costs per task up to 70% lower. Those comparisons appear on Inception's website without a standardized testing methodology or named baseline for every claim, so they remain Inception's measurements. The listed API price for both Mercury 2 and Mercury Edit 2 is $0.25 per million input tokens and $0.75 per million output tokens.

The sales targets reveal where Inception expects parallel decoding to matter first: coding tools, voice agents, search, business automation and creative software. Each involves repeated model calls or a human waiting for an answer. A modest quality improvement can disappear inside a slow interface; a faster model can make an existing capability usable in a live workflow.

Investors have financed that bet. On November 6, 2025, Inception disclosed a $50 million seed round led by Menlo Ventures, with Mayfield, Innovation Endeavors, Microsoft's M12, Snowflake Ventures, Databricks Investment and NVIDIA's NVentures participating. Andrew Ng and Andrej Karpathy joined as angel investors, according to TechCrunch's report on the round.

The round gave Inception the resources to compete with labs capable of testing the same architecture at much larger scale. It also raised the burden of proof. Fast decoding is commercially useful only when the model remains accurate enough for the task and keeps its advantage under real serving loads.

Google and NVIDIA validate the architecture, with caveats

The diffusion approach is no longer Inception's private contrarian bet. Google introduced DiffusionGemma on June 10 as an experimental, open model with 26 billion total parameters and 3.8 billion active parameters during inference. Google claims up to four times faster text generation on dedicated GPUs.

Google also states that DiffusionGemma's overall output quality is below its standard Gemma 4 models and recommends the autoregressive line when maximum quality matters. That admission captures the central commercial tension in diffusion LLMs: parallel generation can lower latency, but developers still have to decide how much model quality they are willing to trade for speed.

NVIDIA's Nemotron-Labs-Diffusion takes a hybrid approach, combining autoregressive, diffusion and self-speculative decoding in one architecture. In diffusion mode, the design can draft tokens in parallel while autoregressive decoding verifies them.

Kuleshov's guide makes room for that uncertainty. It presents diffusion as a collection of methods that can be combined with autoregressive generation, encoder-decoder architectures and post-training techniques. The strongest near-term products may use diffusion selectively, applying parallel refinement where latency or output constraints justify the added complexity.

The final argument reaches beyond faster autocomplete. Kuleshov and his co-authors contend that parallel inference could allow AI labs to spend more useful computation per second during post-training and inference-time reasoning. Transformers made large-scale training practical by replacing sequential recurrent computation with parallel operations. Inception is betting that diffusion can bring a similar hardware fit to generation.

The analogy remains a research thesis, and diffusion models have yet to match the scale, deployment history and independent evaluation available for autoregressive LLMs. The guide's contribution is narrower and immediately useful: it shows builders which mechanisms now exist, how they fit together and where the architecture still needs work. For Kuleshov, that educational work also serves a founder's purpose. A new model category becomes easier to sell once engineers can see how to build it.