Skip to main content
AI Market Watch
Loading...

Two Delta

Category: AI Infrastructure

Stealth Israeli AI lab building technology to accelerate AI model inference and make models use compute more efficiently. Two Delta was founded in 2025. The company is led by Dvir Zeltzer. Based in Israel. Team size: 1-10. Latest round: Seed. Key investors include Bill Ackman, Eric Schmidt.

Founded
2025
Headquarters
Israel
Team size
1-10

Value proposition

Faster, cheaper LLM inference without quality loss. Its DCOCO speculative-decoding method reported 2.03x throughput over standard autoregressive decoding on NVIDIA Nemotron 3.5 Super. These are company-run tests under specified settings.

Products and solutions

DCOCO (DFlash Contextual Corrector) speculative-decoding method, with open weights on HuggingFace (TD-Labs/Nemotron-3.5-Super-VL-DCOCO, TD-Labs/Qwen3-4B-DCOCO) and code at github.com/two-delta/DCOCO. EvoWeb, a benchmark for AI coding agents that build and maintain web apps over 10 successive requests.

Unique value

Improves speculative-decoding drafts by combining candidate-token embeddings with drafter hidden-state context and scoring neighboring-token combinations in parallel. This raises accepted tokens per verification step while preserving target-model output quality.

Target customer

Frontier model developers, inference providers and enterprises running LLMs at scale (company is in stealth; go-to-market not disclosed)

Industries served

AI infrastructure / LLM inference, AI coding agents evaluation

Technology advantage

Deep AI research pedigree via Dr. Eli David (25+ years in neural nets, prior AI exits). Results achieved in collaboration with NVIDIA and benchmarked on DGX B200 with vLLM: 2.03x vs vanilla decoding and 19.8% over DFlash, with gains holding up to 32 concurrent requests. All results are company-reported.

How they differentiate

Contextual corrector layered on parallel drafting (DFlash) that selects more consistent token sequences before verification. Reported 19.8% higher average throughput than DFlash under the same configuration.

Main competitors

Z Lab DFlash (block-diffusion speculative decoding), EAGLE / P-EAGLE (speculative decoding used in vLLM/AWS), inference-optimization providers such as Baseten

Key partnerships

NVIDIA (research collaboration accelerating Nemotron 3.5 Super VL inference with DCOCO)

Major milestones

2025: founded. Oct 2026: published DCOCO results with NVIDIA (2.03x throughput on Nemotron 3.5 Super, company-reported) and the EvoWeb agent benchmark. Oct 2026: seed round of tens of millions of dollars reported by Calcalist. Eli David named in TLV AI 50.

Growth metrics

Fewer than 10 employees (Oct 2026)

Market positioning

Secretive, research-led, very small-team AI lab with high-profile angel backing, focused on AI efficiency and inference acceleration

Geographic focus

Israel-based; global AI infrastructure market

About Dvir Zeltzer

Ex-Chief Finance & Strategy Officer at Wing (Alphabet); ex-Head of Finance at X, the moonshot factory; ex-Google Partnerships Finance Lead. Business and management background, including a senior role at Google.

Latest news about Two Delta

More AI Infrastructure companies

Official website: