Harshit Varma

Inception

@hrshtv

PhD student @UCBerkeley | Prev: Founding Engineer @_inception_ai, Researcher @GoogleDeepMind, CS undergrad @IITBombay

Berkeley, CA hrshtv.github.io Joined July 2021

638 Following 437 Followers 121 Posts

Latest posts 5

Pinned post

Harshit Varma @hrshtv Aug 21

Life update: after nearly 2 incredible years at @_inception_ai, I’m starting a PhD at @UCBerkeley this fall.

It’s been an amazing ride helping build Mercury, Mercury 2, and more from the ground up as a founding member. Huge thanks to @adityagrover_, @volokuleshov, @StefanoErmon, and everyone at Inception for making it so special.

Lately, I find myself excited about topics in self improvement, long-horizon discovery, and multi-agent systems.

Hit me up if you’re around 🐻

Harshit Varma @hrshtv 1d

Excited to present our new work on distillation detection: Subliminal Watermarks (SubWM). SubWM is much more robust compared to existing distillation detectors, is able to distinguish different watermarking keys, and also works reliably for out-of-domain prompts!
Blog: subliminal-watermarks.github.io
🧵

0:24

Harshit Varma @hrshtv Oct 15

Mercury saves you time, take @TIME's word for it

Inception Verified account @_inception_ai Oct 15

Mercury 2 has been named one of @TIME's Best Inventions of 2026!

Our first reasoning diffusion LLM, and the first of its kind in production, was recognized in the AI category.

Thank you to the judges for recognizing diffusion LLMs as one of the technologies shaping the future of AI. The lower latencies offered by diffusion LLMs opens the door for real-time AI systems like coding models, voice agents, and search engines.

It’s a great win for diffusion LLMs. More to come! 🧵

View quoted post on X

Harshit Varma @hrshtv Feb 26

#1/134

Inception Verified account @_inception_ai Feb 26

#1 speed 😎

cc @ArtificialAnlys

View quoted post on X

Harshit Varma @hrshtv Feb 25

Artificial Analysis Verified account @ArtificialAnlys Feb 24

Inception Labs has launched Mercury 2, their next generation production-ready Diffusion LLM. Mercury 2 achieves >1,000 output tokens/s with significant gains in intelligence

@_inception_ai's Diffusion LLMs (“dLLMs”) use a different architecture compared to autoregressive based LLMs. The Diffusion LLM generation process starts with noise and iteratively refines the output using a transformer model that can modify multiple tokens in parallel. This allows parallelization of output token generation, allowing faster output speeds because many output tokens are generated at the same time.

Key takeaways:
➤ Amongst comparable size/price-class models, Mercury 2 performs competitively in intelligence vs. output speed. While it does not have leading intelligence, it’s output speed is more than 3X the next fastest model in this class (benchmarks based on first party endpoints or the median of providers serving the model where a first party endpoint is not available)
➤ Key strengths include agentic coding & terminal use and instruction following. Mercury 2 performs at similar level to Claude 4.5 Haiku on Terminal-Bench Hard and scores 70% on IFBench (Instruction Following), outperforming gpt-oss-120B, GPT-5.1 Codex mini, and GPT-5 nano

Inception Labs background:
This is the second release from Inception Labs. The founders were previously professors from Stanford, UCLA, and Cornell and have contributed to AI research & technologies including Flash Attention, Decision Transformers, and Direct Preference Optimization (DPO).

See below for further analysis.

View quoted post on X