Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Today we're excited to announce Mercury 2.5 It’s the most capable diffusion LLM on the market. It is a 40% jump in intelligence over Mercury 2 and runs at over 1,100 tokens/sec on widely available NVIDIA AI GPUs.

Stefano Ermon

244,887 subscribers

143,813 views • 16 days ago •via X (Twitter)

Science & Technology Education

Anya Rossi• Live Now

Private livecam show

41 Comments

Stefano Ermon16 days ago

Mercury powers real-time systems with the tightest latency budgets. @OpenCall_AI uses Mercury to run voice agents on live patient calls. After switching from an AI inference chip provider, p50 latency fell below 200 ms and p99 fell from several minutes to one second. This keeps multi-step reasoning inside the latency budget of a live call.

Stefano Ermon16 days ago

Mercury 2.5 is available through our API and on @OpenRouter and @Baseten. New accounts include 100M free tokens. At launch, Mercury 2.5 is 80% off at $0.04/M input and $0.15/M output. Try Mercury 2.5:

NVIDIA AI16 days ago

Congrats to the team Stefano!

Deedy16 days ago

@NVIDIAAI Mercury 2.5 continues to redraw the performance-throughput pareto frontier for LLMs

Stefano Ermon15 days ago

@NVIDIAAI Thanks Deedy!

Tim Tully16 days ago

@NVIDIAAI Congrats on the launch! Excited for developers to get their hands on Mercury 2.5 and see what they build.

Barr Yaron16 days ago

@NVIDIAAI let's goooo!!! congratulations @StefanoErmon @volokuleshov @adityagrover_ and team Inception!!!

Stefano Ermon16 days ago

@NVIDIAAI @volokuleshov @adityagrover_ Thanks!

Gabriel Moncha16 days ago

@NVIDIAAI i love the idea of diffusion LLMs but i just wish you've benchmarked it against popular modals too and also mention which TerminalBench version you've tested

Samar Khanna16 days ago

@NVIDIAAI ⚡️

Jeremy Irvin16 days ago

@NVIDIAAI 🧠⚡️📈

Neal Jean16 days ago

@NVIDIAAI let's goooo!! best team 😄

Stefano Ermon16 days ago

@NVIDIAAI Thanks Neal!

Jessica the Recruiter16 days ago

@NVIDIAAI And we’re hiring 😏

Emily Liu16 days ago

@NVIDIAAI 🧨🧨🧨

AGI House16 days ago

@NVIDIAAI Congrats to the whole Inception team on Mercury 2.5! Great to see diffusion models keep proving out in production.

Azalia Mirhoseini15 days ago

@NVIDIAAI This is great, congrats!

Stefano Ermon15 days ago

@NVIDIAAI Thank you Azalia!

Davis Treybig16 days ago

@NVIDIAAI So fast!!

Dídac16 days ago

@NVIDIAAI Congrats on the launch!

Guillem Garcia16 days ago

@NVIDIAAI omg I'm so excited to try this out, thank you!

Stefano Ermon16 days ago

@NVIDIAAI Thank you!

Jaya Gupta16 days ago

@NVIDIAAI Wow this is speedy!!

Tim Shi16 days ago

@NVIDIAAI congrats on the launch @StefanoErmon @volokuleshov !

Diana Vicinanza16 days ago

@NVIDIAAI Congrats on the launch, team! Excited to see what gets built with Mercury 2.5 🚀

Lucas15 days ago

@NVIDIAAI very fast

Stefano Ermon15 days ago

@NVIDIAAI Thank you!

RohaanXD16 days ago

@NVIDIAAI What's the 40% measured on? Genuinely asking the speed numbers are easy to verify but "intelligence" needs an index name to mean anything. Happy to be pointed at it.

Nancy Xu15 days ago

@NVIDIAAI Lets go! Congrats on the launch @StefanoErmon @volokuleshov and team @_inception_ai !

⭕ Slt （苏柏凯）16 days ago

@NVIDIAAI This is blazing fast. congrats.

pikachu guy16 days ago

@NVIDIAAI wow sick

Krrish Agarwalla16 days ago

@NVIDIAAI I am really hyped up for this. Building one myself right now.

vibebuilder16 days ago

@NVIDIAAI HELLLL YAHHHH upgrading my app

Steve16 days ago

@NVIDIAAI excited for diffusion language models.

Erfie16 days ago

@NVIDIAAI congrats

Chieh-Hsin (Jesse) Lai16 days ago

@NVIDIAAI Congrats!

Nando de Freitas14 days ago

@NVIDIAAI Congratulations @StefanoErmon — great to see progress in text diffusion at scale

Neo Vector14 days ago

@NVIDIAAI Reading this made me realize I don't know shit about how diffusion decoding actually works. Guess I have homework this weekend.

T. Rebentisch15 days ago

@NVIDIAAI Comparable quality to: Haiku and Gemini Flash-light 😭 But im glad someone at least continues work on different architectures.

CatGod15 days ago

@NVIDIAAI Congrats! Really impressive to see the team keep shipping Mercury 2.5

Molei Tao14 days ago

@NVIDIAAI Congrats!

Related Videos

Today's autoregressive models generate one token at a time. Mercury 2 generates tokens in parallel. Over 1,000 tok/sec on standard GPUs, at comparable quality to speed-optimized models. Since launch, the community has been showing what diffusion LLMs can unlock. Thanks to the team at Clyep for the breakdown.

Today's autoregressive models generate one token at a time. Mercury 2 generates tokens in parallel. Over 1,000 tok/sec on standard GPUs, at comparable quality to speed-optimized models. Since launch, the community has been showing what diffusion LLMs can unlock. Thanks to the team at Clyep for the breakdown.

Inception

21,231 views • 4 months ago

What if language models didn't have to generate one token at a time? Our CEO Stefano Ermon joined TBPN to break down how Mercury 2's diffusion LLM hits 1,000+ tok/s on standard NVIDIA GPUs — and why speed changes the product for coding, voice agents, and search.

What if language models didn't have to generate one token at a time? Our CEO Stefano Ermon joined TBPN to break down how Mercury 2's diffusion LLM hits 1,000+ tok/s on standard NVIDIA GPUs — and why speed changes the product for coding, voice agents, and search.

Inception

27,188 views • 6 months ago

We're excited to announce Autotab 1.0. The first AI agent that you onboard, not integrate. Models have the intelligence required for most work—but they’re limited by context, tools, and coherence over long tasks. We’ve built Autotab to unhobble models on these bottlenecks 🧵

We're excited to announce Autotab 1.0. The first AI agent that you onboard, not integrate. Models have the intelligence required for most work—but they’re limited by context, tools, and coherence over long tasks. We’ve built Autotab to unhobble models on these bottlenecks 🧵

Autotab

72,945 views • 1 year ago

Top designer at Mercury shows her agentic design explorations in Figma. Designers working on AI features should copy this and sketch how the LLM works before prototyping. 👀 Full conversation on Sneak Peek: carol shows her thought process for designing Mercury Command in Figma👇

Top designer at Mercury shows her agentic design explorations in Figma. Designers working on AI features should copy this and sketch how the LLM works before prototyping. 👀 Full conversation on Sneak Peek: carol shows her thought process for designing Mercury Command in Figma👇

Jayneil Dalal

20,539 views • 2 months ago

🔥BREAKING NEWS! COLONY 🔺 is excited to announce that it successfully acquired over 500,000 $AVAX in the bear market, seizing opportunities. Today marks a pivotal moment—we're launching our Validator Program & unveiling the Sustainability Program. Explore their significance🧵

🔥BREAKING NEWS! COLONY 🔺 is excited to announce that it successfully acquired over 500,000 $AVAX in the bear market, seizing opportunities. Today marks a pivotal moment—we're launching our Validator Program & unveiling the Sustainability Program. Explore their significance🧵

COLONY 🔺

158,482 views • 2 years ago

Mercury Agent is now available on Atomic Bot! Say it once and every session after already has your context. The longer you work together, the less you have to explain. One click and Mercury is live 24/7 in the cloud!

Mercury Agent is now available on Atomic Bot! Say it once and every session after already has your context. The longer you work together, the less you have to explain. One click and Mercury is live 24/7 in the cloud!

atomicbot.ai

26,745 views • 20 days ago

Fast tokens are the most valuable tokens. Because inference speed is productivity. And Cerebras makes the fastest tokens in the world. Google DeepMind's Gemma 4 runs at 1,500+ tokens per second on Cerebras — more than an order of magnitude faster than on GPUs and 15x faster than Claude Haiku 4.5. Private preview today. GA later this month. Build blazing fast multimodal apps.

Fast tokens are the most valuable tokens. Because inference speed is productivity. And Cerebras makes the fastest tokens in the world. Google DeepMind's Gemma 4 runs at 1,500+ tokens per second on Cerebras — more than an order of magnitude faster than on GPUs and 15x faster than Claude Haiku 4.5. Private preview today. GA later this month. Build blazing fast multimodal apps.

Andrew Feldman

12,008 views • 3 months ago

The "is AI UGC good enough?" debate is officially over. Excited to launch Actor+ Most realistic AI UGC library powered by Seedance 2.5 Live today for all Arcads users

The "is AI UGC good enough?" debate is officially over. Excited to launch Actor+ Most realistic AI UGC library powered by Seedance 2.5 Live today for all Arcads users

Romain Torres

18,570 views • 1 month ago

Physical AI is the biggest opportunity most people are sleepwalking past. "Physical AI is a $50 trillion market." - Jensen at The All-In Podcast Robots. Autonomous vehicles. Embodied AI. The infrastructure to power all of it. Nvidia isn't just selling GPUs anymore. They're building the OS for the physical AI era.

Physical AI is the biggest opportunity most people are sleepwalking past. "Physical AI is a $50 trillion market." - Jensen at The All-In Podcast Robots. Autonomous vehicles. Embodied AI. The infrastructure to power all of it. Nvidia isn't just selling GPUs anymore. They're building the OS for the physical AI era.

Shruti

51,545 views • 6 months ago

Our internal AI factory serves 4 trillion tokens a month to NVIDIA employees, and demand is still growing 40% month over month. In Episode 2 of AI Factory Insider, we get into what it actually takes to run that at scale: the architecture, the real use cases, and what we learned along the way. Catch the full episode ➡️

Our internal AI factory serves 4 trillion tokens a month to NVIDIA employees, and demand is still growing 40% month over month. In Episode 2 of AI Factory Insider, we get into what it actually takes to run that at scale: the architecture, the real use cases, and what we learned along the way. Catch the full episode ➡️

NVIDIA AI Infrastructure

20,714 views • 2 months ago

Sec. Kennedy: Mercury is absolutely—it’s the most powerful neurotoxin we know of in the universe. Rep. Harder: So why are you working for an administration that is doubling the amount of mercury pollution in our air and water? Sec. Kennedy: I am running HHS and I’m ending the chronic disease epidemic. I’m not going to comment on that because I don’t know anything about it. Rep. Harder: Well, that’s very convenient for you, and it shows a significant degree of cowardice. It’s not just mercury. President Trump in July 2025 granted 50 chemical plants— Sec. Kennedy: Why don’t you ask me something that I’m in charge of instead of other things? Rep. Harder: Are you not in charge of Health and Human Services?

Sec. Kennedy: Mercury is absolutely—it’s the most powerful neurotoxin we know of in the universe. Rep. Harder: So why are you working for an administration that is doubling the amount of mercury pollution in our air and water? Sec. Kennedy: I am running HHS and I’m ending the chronic disease epidemic. I’m not going to comment on that because I don’t know anything about it. Rep. Harder: Well, that’s very convenient for you, and it shows a significant degree of cowardice. It’s not just mercury. President Trump in July 2025 granted 50 chemical plants— Sec. Kennedy: Why don’t you ask me something that I’m in charge of instead of other things? Rep. Harder: Are you not in charge of Health and Human Services?

Acyn

3,556,097 views • 5 months ago

Excited to share what my team has been working on lately - Gemini diffusion! We bring diffusion to language modeling, yielding more power and blazing speeds! 🚀🚀🚀 Gemini diffusion is especially strong at coding. In this example the model generates at 2000 tokens/sec, including overheads like tokenization, prefill, safety filters etc.

Excited to share what my team has been working on lately - Gemini diffusion! We bring diffusion to language modeling, yielding more power and blazing speeds! 🚀🚀🚀 Gemini diffusion is especially strong at coding. In this example the model generates at 2000 tokens/sec, including overheads like tokenization, prefill, safety filters etc.

Brendan O'Donoghue

578,018 views • 1 year ago

On the periodic table, nothing separates gold (Au) from mercury (Hg); they are direct neighbors in the sixth period. Gold has the atomic number 79 (Group 11), and mercury has the atomic number 80 (Group 12). When these two elements mix, an immediate physicochemical reaction occurs in which the mercury dissolves the gold, forming a metallic alloy called an amalgam. The process is fully reversible and the gold is not lost, though the jewelry requires chemical or thermal treatment to return to its original state. Mercury does not destroy gold; it merely mixes with it at the surface. Since mercury evaporates at a much lower temperature than the melting point of gold, it can be completely removed. There was once a "legend" that the ancients could turn mercury into gold, though "EXPERTS" deny this. Today, however, through the use of nuclear physics, such a transformation is possible by altering the heart of the atom itself. Since mercury has 80 protons in its nucleus and gold has 79, modern science can "pluck" a proton from mercury to transform it into genuine gold.

On the periodic table, nothing separates gold (Au) from mercury (Hg); they are direct neighbors in the sixth period. Gold has the atomic number 79 (Group 11), and mercury has the atomic number 80 (Group 12). When these two elements mix, an immediate physicochemical reaction occurs in which the mercury dissolves the gold, forming a metallic alloy called an amalgam. The process is fully reversible and the gold is not lost, though the jewelry requires chemical or thermal treatment to return to its original state. Mercury does not destroy gold; it merely mixes with it at the surface. Since mercury evaporates at a much lower temperature than the melting point of gold, it can be completely removed. There was once a "legend" that the ancients could turn mercury into gold, though "EXPERTS" deny this. Today, however, through the use of nuclear physics, such a transformation is possible by altering the heart of the atom itself. Since mercury has 80 protons in its nucleus and gold has 79, modern science can "pluck" a proton from mercury to transform it into genuine gold.

Bronze Giant

491,675 views • 12 days ago

Mercury 2 is live 🚀🚀 The world’s first reasoning diffusion LLM, delivering 5x faster performance than leading speed-optimized LLMs. Watching the team turn years of research into a real product never gets old, and I’m incredibly proud of what we’ve built. We’re just getting started on what diffusion can do for language.

Mercury 2 is live 🚀🚀 The world’s first reasoning diffusion LLM, delivering 5x faster performance than leading speed-optimized LLMs. Watching the team turn years of research into a real product never gets old, and I’m incredibly proud of what we’ve built. We’re just getting started on what diffusion can do for language.

Stefano Ermon

1,049,004 views • 7 months ago

Gemma 4 31B is now available in Public Preview on Cerebras. Our first multimodal model runs at over 1,800 tokens/s for ultra-fast image and text workflows. Give it a try:

Gemma 4 31B is now available in Public Preview on Cerebras. Our first multimodal model runs at over 1,800 tokens/s for ultra-fast image and text workflows. Give it a try:

Cerebras

282,043 views • 2 months ago

We're excited to announce Glympse is migrating to . We’ve partnered with Time to adopt their creator tokens and build directly on top of their creator market. Now, one token captures it all. Your time. Your attention. Tokenized together. Speculate on access and engagement in a single market. Creator capital markets supercharged. Here's how it works 👇

We're excited to announce Glympse is migrating to . We’ve partnered with Time to adopt their creator tokens and build directly on top of their creator market. Now, one token captures it all. Your time. Your attention. Tokenized together. Speculate on access and engagement in a single market. Creator capital markets supercharged. Here's how it works 👇

glympse.fun

147,893 views • 1 year ago

Today we're releasing Not Diamond… The world’s most powerful AI model router. Not Diamond maximizes LLM output quality by automatically recommending the best LLM on every request at lower cost and latency. And it takes <5m to set up. Watch this to see how to start using it:

Today we're releasing Not Diamond… The world’s most powerful AI model router. Not Diamond maximizes LLM output quality by automatically recommending the best LLM on every request at lower cost and latency. And it takes <5m to set up. Watch this to see how to start using it:

Tomas Hernando Kofman

76,490 views • 2 years ago