x
The artificial intelligence race has officially entered a new, blistering phase. For years, interacting with large language models has felt a bit like watching a ghost type on an old-school typewriter. The model generates one word, looks at what it just did, predicts the next word, and repeats the process until the sentence is finished. It is effective, but inherently bottlenecked.
Now, a major architectural shift is under way, and a new player has just thrown down the gauntlet. Inception Labs has officially unveiled Mercury 2, a reasoning language model that completely abandons the traditional typewriter approach. Generating an astonishing 1,000 tokens per second, Mercury 2 has managed to beat Google’s new DiffusionGemma at its own game, proving that ultra-high-speed AI does not have to come at the cost of intelligence.
To understand how Mercury 2 achieves these speeds, we have to look under the hood at an approach known as diffusion. Traditional text chatbots are autoregressive, meaning they generate text sequentially. Diffusion models, however, work more like the AI image generators we use to create digital art.
Instead of writing word by word, a diffusion text model fills an entire block of text with random placeholder tokens, effectively creating a wall of digital static. It then erases this noise across a handful of parallel passes, refining the entire block of text simultaneously until a coherent, finished response locks into place all at once.
The result is a monumental leap in throughput. While premium sequential models like Anthropic’s Claude Haiku 4.5 Reasoning clock in at roughly 89 tokens per second, and OpenAI’s GPT-5 Mini hits around 71, Mercury 2 leaves them in the dust at the 1,000 tokens per second mark.
Google recently attempted the exact same trick with its open-weight model, DiffusionGemma, hitting similar blistering speeds. However, dropping the typewriter approach usually comes with a penalty: a noticeable drop in the model's analytical capabilities. This is where Mercury 2 has pulled off something of a miracle, maintaining high intelligence while running at hyper-speed.
The divergence between the two models becomes clear when looking at the AIME 2026 benchmark, a rigorous test built from real American Invitational Mathematics Examination problems.
To put that into perspective, Google's standard, non-diffusion Gemma 4 scores 88.3% on the same test. Google's own developer guide even concedes the quality drop, recommending standard Gemma 4 over DiffusionGemma for tasks requiring maximum reasoning quality. Mercury 2, on the other hand, beats Google’s flagship standard model while running at a fraction of the time. On the GPQA benchmark—a PhD-level science test—the two diffusion models performed much closer, with Mercury 2 leading at 77% over DiffusionGemma's 73.2%.
Welcome to the diffusion era.
— Inception (@_inception_ai) June 18, 2026
We bet on parallel generation years ago, when it was a contrarian idea. It's great to see the industry arrive.
Mercury 2 continues to lead the Pareto frontier for quality, speed, and cost among publicly available diffusion LLMs. pic.twitter.com/qSHuiR7vmH
These benchmarks are translating directly into massive efficiency gains in the wild. Augment Code, an AI coding-agent platform, recently swapped out Anthropic’s powerhouse Claude Opus 4.7 in favour of Mercury 2 for its context-compaction subagent. The move resulted in an 82% drop in latency and a massive 90% cut in costs, all while maintaining the exact same output quality.
This revolutionary architecture is built on the academic foundations of Inception Labs' founder, Stefano Ermon. As a Stanford professor, Ermon co-authored some of the foundational score-based diffusion techniques that power modern image generators. The startup's potential has already turned heads in Silicon Valley, securing a 50 million dollar funding round backed by Nvidia’s venture arm, alongside prominent AI pioneers Andrew Ng and Andrej Karpathy.
For everyday users and digital creators, the real magic of this breakthrough is something you feel rather than read on a spec sheet: the "flow."
When you are deep in a brainstorming session, writing complex code, or building an intricate plan, waiting for a traditional AI model to finish its thoughts breaks your cognitive momentum. Hyper-fast diffusion models make the AI feel like it is keeping perfect pace with your brain. It allows for instant auto-completion, rapid-fire iterations, and voice interfaces that respond without a hint of awkward lag.
More importantly, this speed completely unlocks the next generation of AI development: multi-agent ecosystems. Modern AI systems are no longer just one massive, slow model doing all the heavy lifting. Instead, they operate like an orchestra of specialised helpers. You might have one subagent for deep reasoning, three for quick summaries, another for looking up tools, and one more for checking the final output.
When these subagents rely on traditional sequential models, making so many background calls becomes incredibly slow and prohibitively expensive. Parallel diffusion models make these background utility calls so cheap and fast that developers can use them liberally, leading to far more autonomous and capable AI assistants.
Video - OpenAI vs Inception
While Mercury 2 is a massive leap forward, there are a few realities to keep in mind. For the absolute hardest frontier reasoning problems, the largest, heaviest autoregressive models may still hold a slight edge for now. Furthermore, unlike Google’s DiffusionGemma, which is free and open-weight on Hugging Face, Mercury 2 is a paid, closed-weight API model. The broader developer ecosystem of local runtimes and agent frameworks will also need a bit of time to catch up before this parallel architecture becomes seamless everywhere.
Nevertheless, by proving that text diffusion can be lightning-fast without losing its smarts, Inception Labs has redefined what is possible on standard, commodity GPUs. The era of the AI typewriter is officially coming to a close.
To read the original report and find out more details, you can view the full article on Decrypt:
👉 Inception Labs' Mercury 2 AI Beats Google's DiffusionGemma at Its Own Game
Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.
