Volodymyr Kuleshov 🇺🇦

Really excited for Voyager, Anvisha and team are on fire!!

We're excited to welcome Michael Giannangeli (@mrgiannangeli) to Inception as our new Head of Product! 📷

Michael has spent years building AI products that millions of people use every day. At Amazon, he led product and agentic AI for Nova foundation models, building the capabilities for intelligent systems to reason, act, and scale in the real world. Before that, Michael ran the Echo and Echo Dot product lines, started Alexa’s computer vision team, and helped scale Alexa from 1M to 100M+ customers.

He holds an MBA from MIT. Before tech, he worked as an economic consultant, did stints on Wall Street and Capitol Hill, and played NCAA basketball.

At Inception, Michael will lead product as we bring Mercury, our family of diffusion LLMs, to more developers and enterprises. Mercury generates tokens in parallel rather than one at a time, so it's much faster and cheaper than standard autoregressive models. Michael knows how to turn breakthrough AI into products people love, which makes him a great fit for this next phase.

Welcome to the team, Michael! 🎉

Welcome, Stefan!

Great overview by Stefan from our team on System 1 vs. System 2 thinking in AI models + how the Mercury Decide API fits in this picture 👀

🎙️ Announcing Mercury Voice, the newest model in the Mercury lineup

⚡ Powered by diffusion, it's the world's fastest LLM for real-time voice agents in customer support, sales, education & more

🏎️ Faster than Cerebras
🧠 Higher quality than frontier models

Give it a spin!

📢 Mercury 2.5 is here!

🚀 Since Mercury 2, usage has grown over an order of magnitude, with customers running latency-sensitive workloads across search, voice and coding agents.

Now: 40% more intelligence than Mercury 2, 1,100+ tok/s, lower price, 260K context + better agent tooling.

Check out the full thread for all benchmarks and results!

🚀 Mercury 2.5 is here — the next generation of Mercury diffusion models.

Big gains on agentic workloads, still running at >1000 tok/sec in production.

Live today on OpenRouter — check it out 👇

🙏

Thanks for having me! It was super fun talking about Mercury diffusion LLMs at YC last night.

Our work on Mercury 2 diffusion language models is covered by the MIT Tech Review as one of four exciting research directions where LLMs are headed.

Check out the story here: technologyreview.com/2026/08/10/114…

I'm super excited about this new workshop dedicated to diffusion language models at Neurips2026!

If you are working on diffusion LMs, this is a great venue to submit your work. I'm looking forward to being there in Sydney this winter!

“Noon, Leon, biology professor, is playing tennis. Qwen3.8 is verifying protein sources.”

Love it

Excited for Mercury 2 to be launching on Baseten Model Labs today!

This launch makes Mercury diffusion LLMs available to an even larger set of enterprise customers with ultra-high speed, compliance, auto-scaling, pay-as-you-go pricing, and more.

To celebrate, we are increasing by 10x the free token tier of Mercury 2 on Baseten. Give it a spin!

Where do diffusion language models shine? One major use case is voice agents.

With diffusion LLMs running at >1000 tok/sec, the model can reason in real time before producing outputs. This boosts quality, and latency is still low enough for a live phone call.

Check it out 👇

This Friday at 13:30-2:30pm I'm moderating a panel on autoregressive vs. diffusion language models at the SPIGM workshop in Hall D1 at #ICML2026.

The lineup of panelists is really exciting, come join us!

New blog post: How to Build a Diffusion Language Model.

Diffusion LLMs went from open problem to reality in 2 years (Mercury, Gemma Diffusion, Nemotron Diffusion). With my Cornell group, we wrote up the research advances that make them work.

We use today's OSS large diffusion models as examples and show how they're built from core techniques:
• masked diffusion (MDLM)
• iterative refinement (UDLM, ReMDM)
• variable-length generation (block diffusion, encoder-decoder architectures)
• controllable generation (D-CFG, D-CBG)
• fast samplers (Duo)
• RL post-training (d1, d2)

This post covers research led by a talented group of Cornell PhD students including @mariannearr @SchiffYair @Guanghan__Wang

The content is adapted from talks I gave this year on dLLMs. It covers most of the main ingredients behind open-source diffusion language models today.

Link: kuleshov-group.github.io/blog/blog/2026…

I'm at #ICML2026 this week, come talk to me about all things diffusion!

On Wed morning, my student @Guanghan__Wang is presenting d2, improved techniques for RL in diffusion LLMs.

On Thu morning, @mariannearr is presenting Set Diffusion, the next generation of block diffusion supporting any-order generation.

On Fri, I am moderating a panel on diffusion vs. autoregression at the SPIGM workshop with an exciting set of speakers.

Hope to see you in Seoul! Links to @Guanghan__Wang @mariannearr papers are below.

Super exciting that another ICML best paper award goes to work on diffusion language models.

That's now three ICML's in a row where one of the best papers is on diffusion for language: SEDD (2024), Train for the Worst and Plan for the Best (2025), and now the Flexibility Trap (2026).

Last week, I had the pleasure to give a talk at the 2026 ML Summer School at @Columbia.

My lecture was on recent work on diffusion language models from my group at Cornell and at Inception, including MDLM, UDLM, BD3LM, ReMDM, E2D2, and work in industry on scaling up these methods for modeling language and biological sequences.

Slides and papers are linked below. 🧵

Slides: drive.google.com/file/d/1JXUGE_…

Had a blast giving a talk on diffusion language models to an awesome group of PhD students at MLSS 2026 last week!

Thank you for inviting me and thanks for all the great questions 🙏