Twitxy Saved

Pavan Katta

Verified account This account is verified. Learn more

@PavanKatta1998

Co-founder @hopper_research (YC F26). voice inference
San Francisco pavankatta.com Joined September 2023
764 Following 613 Followers 650 Posts

Latest posts 5

Pavan Katta Verified account @PavanKatta1998
as much sympathy as i have for teams building models, i’ve completely stopped trusting internal benchmarks any company publishes, including ours.

for founders of these companies, don’t let your research team fool you with made up metrics to justify their release.

seen so many straight up bad models that claim SOTA, it’s not worth even looking at any eval that’s not 3rd party.
6 replies 0 reposts 20 likes 1.1K views View on X
Pavan Katta Verified account @PavanKatta1998
our production screenshot btw. 100ms TTFT and 1000 TPS
Sarvam Verified account @SarvamAI 4d
For real-time conversations, Gemma 4 31B is an efficient model that combines agentic capability with Indian language support.

On Sarvam Token Factory it returns its first token 5× sooner and generates 1.6× more tokens per second than the only other provider serving it.

We are already serving 150 billion tokens per day of Gemma 4 31B.
View quoted post on X
4 replies 0 reposts 52 likes 4.8K views View on X
Pavan Katta Verified account @PavanKatta1998
this is the chart I'm most proud of, over any ARR or growth metrics.

We've worked really hard to make our infra stable and able to hit

P90 TTFT of sub 100ms
P90 TPS of 1000 Tok/s

for Gemma 4 which has 31B active params (higher than Deepseek and GLM 5.3 Flash).

It's really easy to vibecode a kernel or an optimization config that has a good headline result.

But for production system serving millions of requests over a fleet of heterogenous GPUs, old school white boarding is still the way.
5 replies 0 reposts 16 likes 1.2K views View on X
Pavan Katta Verified account @PavanKatta1998
we'll get audio perception 2 out in ~3 months.
Scaling up data by 60x this time.
Yi Zhong 🍊 Verified account @yiz_be_building Nov 12
We are updating Vocal Affect Bench (vocalaffectbench.com) with 8 new models, including a suite of audio native decision models and more omni models from @thinkymachines , @Google , @Alibaba_Qwen , @modulate_ai , @TazoLabs , @NVIDIAAI , and @OpenAI .

The leader is still @hopper_research 's Audio Perception-1, but @thinkymachines 's Inkling Small is the new #2, replacing Oruk Resonance 2.

Congratulations @thinkymachines !
View quoted post on X