Inception co-founder Aditya Grover says Google's DiffusionGemma validates the shift to parallel-generation diffusion LLMs

Mercury vs. Gemma Diffusion

Mercury provides comparable speeds of 1200 Tok/sec, while maintaining state of the art quality

Welcome to the diffusion era.

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.

@aftertherain022 @_inception_ai @StefanoErmon Take a look at this talk

https://youtu.be/r305-aQTaU0?si=KSHgmWlLiAXMTEPc

@_inception_ai I’ve been beating this drum!

Pricing: $0.25 / 1M in and $0.75 / 1M out http://inceptionlabs.ai/models

@_inception_ai @grok what Claude model is comparable in coding benchmarks to Mercury 2? How well does it do with frontend design tasks?

@_inception_ai @StefanoErmon I’ve been interested in diffusion-based language models for a while and have been following them closely. I did a deep dive into Mercury while researching for a capstone project. Really interesting work. I’m not convinced it’ll beat chain-of-thought style reasoning, but who knows

@_inception_ai We need a small open weights Mercury model able to compete with Qwen3.6 27b and Gemma 31B, dense, no vision, just text and tool calling

@_inception_ai Impressive numbers. What batch size are you using to hit that cost?

@_inception_ai this is very interesting

@_inception_ai tps?

@_inception_ai Building an app that’s solely relying on Mercury 2!

@_inception_ai You know that they're on 1m context now.?

Mercury 2's coding benchmarks (LiveCodeBench ~67-74, solid agentic/coding scores) put it roughly on par with **Claude 4.5 Haiku** in quality — both strong fast-tier performers.

For frontend design tasks, Mercury 2 does very well. Its diffusion model excels at parallel generation + rapid iterative refinement, making it outstanding for quick UI prototyping, Tailwind/React components, animations, and tight edit loops. Speed advantage shines here for interactive work.

Claude Sonnet variants may edge it on complex/creative designs, but Mercury 2 wins on responsiveness and cost for most frontend flows.

@_inception_ai @Presidentlin

@_inception_ai Betting on parallel generation when it was contrarian and now seeing the whole industry follow is a good reminder that being early is usually lonely before it's right.

@_inception_ai take yang bagus sih ini

@arthopacini @_inception_ai @StefanoErmon ill have a look on my lunch break cheers :)