What Bagel Labs does

Bagel Labs is a privately held AI research lab founded in 2023 and based in San Francisco that develops Distributed Diffusion Models (DDM), a method for training frontier diffusion models across heterogeneous commodity hardware instead of a single homogeneous GPU supercluster. DDM replaces one large model with an ensemble of smaller expert models trained independently on partitioned data with no gradient synchronization, combined at inference by a lightweight router. The company has publicly released two models built on this architecture: Paris-1, the first public DDM for image generation, which achieved a 24% FID improvement using 14x less data and 16x less compute than prior decentralized baselines; and Paris-2, a video DDM with three 11B experts plus a router that beat a matched-compute monolithic baseline by 50%+ on FVD. A dedicated inference runtime, the Paris Inference Engine (PIE), ensembles expert outputs at inference.

The company is now applying DDM to Physical AI workloads in robotics, autonomy, simulation, and world modeling. The go-to-market motion is early-stage enterprise sales, with the first commercial hire (Head of GTM, Physical AI) building a pipeline from outreach through technical evaluation, design partnership, and paid pilot. Pricing is undisclosed; the company is explicitly pre-revenue with no named customers. Distribution and awareness currently rely on the Everything Bagel Substack (130,000+ subscribers), direct research publication, and a small social presence on X and LinkedIn. The company has raised approximately $8.6 million in aggregate (a $3.1M seed led by CoinFund in January 2024 and a $5.5M round led by Polychain in June 2025) and has a team of fewer than 10 people with three open technical and commercial roles.

Name

Bagel Labs

Legal name

Bagel Labs

Website

https://bagel.com

Company type

Private

Founded year

2023

Operating status

Operating

Headcount range

1–10 employees

Short description

Bagel Labs is an AI research lab building Distributed Diffusion Models, a method that trains frontier diffusion models across commodity hardware via independent expert ensembles. It now targets Physical AI teams in robotics, autonomy, simulation, and world modeling through enterprise pilots.

Ownership category

akta.pro rank

Product category

AI/ML Training Infrastructure

akta.pro primary industry

Model Deployment, Serving & Inference Platforms (HDAAABAF)

akta.pro secondary industry

Prompt Engineering, Orchestration & LLMOps Tooling (HDAAACAD)

Keywords

• Distributed model training
• Diffusion models
• AI research infrastructure
• Physical AI training
• Generative AI models

Headquarters

HQ city

San Francisco

HQ country

United States

HQ region

North America

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Marketing or Sales, Operations

Revenue model

1. Physical AI Training Infrastructure Licensing/Pilot: Early-stage company targeting physical AI teams in robotics, autonomy, simulation, and world modeling. Revenue model through design partnerships and paid pilots. Market barely exists yet; first commercial hire building the pipeline from outreach through evaluation and close.

Go-to-market motion1 record

Distribution channels1 record

Marketing channels4 records

Core offering

Bagel Labs develops Distributed Diffusion Models (DDM), a training methodology that replaces a single large diffusion model with an ensemble of smaller expert models trained independently on partitioned data with no gradient synchronization. It ships the Paris-1 image and Paris-2 video generative models as public demonstrations and offers the Paris Inference Engine (PIE) to run the ensemble at inference, targeting physical AI teams in robotics, autonomy, simulation, and world modeling.

Product overview

Bagel Labs is an AI research lab developing Distributed Diffusion Models (DDM), a novel method for distributed training of frontier diffusion models on commodity hardware. The core product portfolio includes: DDM as the foundational methodology, Paris-1 as the first publicly released image generation DDM, Paris-2 as the video generation DDM with three 11B experts, and PIE (Paris Inference Engine) as the inference runtime that combines expert model outputs via a lightweight router. Together, these products enable training of state-of-the-art generative models for robotics, video, and world modelling across heterogeneous hardware.

Differentiator

Problem solved

Functional benefit

Brands

• DDM (Distributed Diffusion Models): Core technology replacing single large diffusion models with ensembles of smaller expert models trained independently with no gradient synchronization
• Paris-1
• Paris-2
• PIE (Paris Inference Engine)
• Everything Bagel

Products and services

• Distributed Diffusion Models (DDM) Foundational training methodology that trains an ensemble of smaller expert models independently on partitioned data without gradient synchronization, removing the tight coupling that forces conventional training onto homogeneous GPU superclusters. Used by physical AI teams in robotics, autonomy, simulation, and world modeling.
• Paris-1 First publicly released DDM for image generation, demonstrating 24% FID improvement (22.60 vs 29.64) on standard benchmarks using 14x less data and 16x less compute than prior decentralized baselines. For AI researchers and developers building generative image systems.
• Paris-2 Video DDM pre-trained from scratch with three 11B video experts and a lightweight router, achieving 50%+ FVD improvement plus 7.2% CLIP, 2.9% aesthetic, and 28.3% motion gains over a matched-compute monolithic baseline. For teams building video generation and world modeling systems.
• Paris Inference Engine (PIE) Inference runtime that runs the DDM ensemble, combining outputs from multiple expert models via a lightweight router at inference time. Used by customers operating the Paris model family or building similar ensemble inference systems.

Quantifiable outcome

• 24% FID improvement (22.60 vs 29.64) using 14x less data and 16x less compute
• +1 more outcomes

Segments1 record

Ideal customer profiles1 record

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability5 records

Feature3 records

Scale indicators5 records

Recent moves6 records

Expansion highlights5 records

Broad incumbents

• MosaicML (Databricks): Pre-acquisition, MosaicML pioneered efficient large-model training on commodity hardware; now part of Databricks. Comparable as a benchmark for what efficient-training infrastructure can become at scale and a potential incumbent competitor.
• Hugging Face: Hub and infrastructure provider for open generative models, including diffusion. Comparable as a destination where Bagel's Paris-1 and Paris-2 models might be distributed and as a potential enterprise training partner or competitor.

Emerging players

• Lightricks: Generative AI company shipping diffusion-based image and video tools to creators and enterprises. Comparable as a generative AI vendor pursuing commercial diffusion-model products and physical-world video use cases.
• Prime Intellect: Decentralized AI training startup pursuing distributed training of large models across heterogeneous hardware. Closely comparable in research direction (commodity-hardware distributed training) and likely overlapping investor and customer mindshare.
• Nous Research: Open AI research lab focused on fine-tuning and training large models on consumer hardware. Comparable as a small, research-driven lab pursuing compute-efficient training and distribution via open releases.

Direct peers

• Together AI: Operates a distributed AI cloud and training/inference platform purpose-built for large model workloads across heterogeneous hardware. Most directly comparable to Bagel on the distributed-training-infrastructure thesis, though focused primarily on LLMs rather than diffusion models.
• Runway: Applied AI lab building frontier video generation models (Gen series) for creative and physical-world use cases. Comparable as a frontier video diffusion model builder pursuing enterprise and physical-AI-adjacent use cases.
• Midjourney: Closed-source image generation service built on diffusion models. Comparable as a frontier-class diffusion model builder competing for enterprise and creative workloads.
• Anyscale: Commercial steward of Ray, a distributed compute framework widely used for large-model training and serving. Comparable as infrastructure-layer distributed AI compute, though broader than diffusion-only training.
• Stability AI: Developer of open-source diffusion models (Stable Diffusion) and the underlying training infrastructure. Directly comparable as a generative AI research-oriented company producing diffusion models and competing for the same enterprise and developer mindshare.

Market position

Strengths4 records

Weaknesses4 records

Competitive moat3 records

Key risks6 records

Key highlights6 records

Customer concentration

• Xx.com/bageldotcom
• LinkedInlinkedin.com/company/bageldotcom

Revenue estimate

Valuation estimate

Number of profiles

Profiles1 record

Funding overview

Funding rounds2 records

Investors11 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Bagel Labs