Israeli startup Two Delta, specializing in artificial intelligence, has secured a significant seed funding round, amassing tens of millions of dollars. This funding round saw participation from notable investors, including Bill Ackman, the founder and CEO of Pershing Square, and Eric Schmidt, the former chairman of Google, as reported by Calcalist.
Established in 2025, Two Delta operates discreetly, sharing minimal information about its activities beyond academic papers available on its website. The company was co-founded by Dvir Zeltzer and Dr. Eli David, focusing on developing technologies that enhance the speed and efficiency of AI model performance while optimizing computing resource usage.
Dr. Eli David, an experienced AI researcher and entrepreneur, collaborates with Dvir Zeltzer, who brings a solid foundation in business management, including previous leadership at Google.
In October 2026, Two Delta shared findings from a partnership with Nvidia, showcasing its DCOCO technology, which accelerates inference processes for Nvidia’s Nemotron 3.5 Super VL model. This innovative approach aims to boost the number of tokens generated per verification step without sacrificing the quality of the output.
DCOCO enhances a methodology known as speculative decoding, where a quicker, smaller model generates a sequence of tokens for a larger model to validate in a single go. While this can expedite the generation process, it may lead to the rejection of some proposed tokens. The challenge lies in crafting drafts that the larger model is more likely to endorse.
To tackle this, Two Delta’s technique integrates candidate-token embeddings with contextual data derived from the hidden states of the drafting model, utilizing insights from the target model’s activations. This scoring system evaluates pairs of neighboring tokens concurrently, thereby selecting more coherent sequences prior to verification by the larger model.
In tests conducted across ten datasets encompassing mathematics, programming, and general tasks, Two Delta claimed that DCOCO achieved a throughput average of 2.03 times higher than standard autoregressive decoding when reasoning was disabled and greedy decoding applied. In a similar setting, it reported a throughput enhancement of 19.8% over the competing DFlash method. The testing, done on Nvidia’s DGX B200 hardware using the vLLM inference framework, yielded additional throughput improvements across various decoding configurations and up to 32 simultaneous requests.
These reported results stem from Two Delta's own studies rather than independent evaluations. The company asserts that DCOCO enhances token generation efficiency while maintaining the output quality of the target model.
Additionally, Two Delta introduced EvoWeb, a benchmark aimed at assessing the effectiveness of AI coding agents in developing and maintaining web applications amid users' continuous requests for new features and modifications. Unlike traditional benchmarks that focus on single-prompt performance, EvoWeb challenges the AI to manage an application through ten successive requests, beginning with an initial build followed by nine modifications.
During the evaluation of 32 web applications, no model was able to sustain complete functionality throughout the testing process. The highest-scoring models, Fable 5 and Kimi K3, achieved scores of 0.694 and 0.684, respectively, on a scale from zero to one.
The primary challenge identified was the integration of new features with the existing system functionality. Although occurrences of build failures were infrequent, the study noted an increase in runtime errors, which rose from 2% at the outset to 12% after a series of ten requests.
