Reducto (@reductoai) / X

Reducto

569 posts

Reducto

@reductoai

The complete agentic document platform for leading AI teams needing performance at enterprise scale. reducto.ai

San Francisco, CA

Joined January 2024

Affiliates Replies Media

Pinned
Reducto
@reductoai
Oct 14, 2025
We raised $75M in Series B funding led by @a16z, bringing our total funding to $108M. @reductoai is defining the new frontier for document intelligence, helping the world’s leading AI teams unlock data from their unstructured documents. We’re launching a new pay-as-you-go tier
Reducto
@reductoai
Jul 30
We'll be at Ai4 in Vegas next week! Find Reducto at Booth P8 🔮 Our events lineup: 🗓️ Tuesday, August 4th: @aditabrm will be speaking on a panel about leveraging Data Strategy to Enable Effective AI 📍 Titian Ballroom 2302 ⏰ 3:00 - 3:45 PM 🗓️ Wednesday, August 5th: Join us for
Reducto
@reductoai
Jul 28
We trained a multi-token prediction head with Megatron and swept concurrency from 1 to 256 on a 9B model served with vLLM across text, key-value, and table workloads. At low concurrency, MTP delivered up to 52% throughput improvement on tables, but by c=64, every workload was
Jean Billa
@jeanghislainbil
Jul 28
Article
When does accepted speculative decoding translate into higher throughput and lower latency?
Speculative decoding is a technique that allows for a cheap drafter to predict the next n tokens. The drafter proposes several tokens, the target model verifies them in one pass. If enough draft...
Reducto
@reductoai
Jul 20
Back in April, we introduced Deep Extract, an agentic extraction mode built for documents that break every standard pipeline. Since then it has processed more than 120 million fields on documents up to 3000 pages long in production. Deep Extract costs more and takes longer than
Reducto
@reductoai
Jul 14
Mercor's APEX benchmark tests whether frontier models can do the actual work of investment banking associates, management consultants, big law associates, and physicians. Every source document in it was parsed by Reducto. APEX-v1-extended grades 10 frontier models against 400