What Fireworks AI does

Fireworks AI is a privately held AI inference and training platform founded in 2022 in Redwood City, California, by the leadership and core engineering team behind PyTorch at Meta — including CEO Lin Qiao (former Head of PyTorch) and CTO Dmytro Dzhulgakov (PyTorch core maintainer). The company operates a disaggregated inference engine optimized from custom kernels through memory management, with proprietary components including FireAttention (claimed 4x faster than vLLM), speculative decoding for sub-100ms streaming responses, multi-node expert parallelism for trillion-parameter MoE models, Multi-LoRA serving, and KV-cache routing for long-context workloads. The training stack spans supervised fine-tuning, DPO, and reinforcement fine-tuning, with a frontier RL toolchain (GRPO, DAPO) and a Training Agent module that takes a model from data upload to production deployment; trained checkpoints deploy to live endpoints in seconds with published K3 KL divergence parity checks. Following the March 2026 acquisition of Hathora, Fireworks runs across 14 regions, multiple bare-metal providers, and four clouds.

The platform serves a hybrid customer base. On the developer side, an API-first self-serve product delivers pay-per-token serverless inference with OpenAI and Anthropic API compatibility, $1 in free credits at signup, cached input tokens at 50% off, and batch inference at half-price; on-demand dedicated GPU deployments are billed per-second at $7–$12 per GPU-hour across H100, H200, B200, and B300 silicon. Enterprise customers buy reserved capacity and Bring-Your-Own-Cloud deployments under multi-year contracts with a 99.9% SLA. The company distributes through Microsoft Foundry (general availability announced at Build 2026), the MongoDB for Startups program, accelerator partnerships with Y Combinator, AWS Activate, and Google for Startups, and a direct field-sales team for Fortune 500 procurement.

Fireworks serves four named customer segments: AI-native startups (Cursor, Vercel, Genspark, Factory AI, Cognition, Lovable), enterprise companies (Samsung, Uber, Doordash, HubSpot, UiPath, GitLab, Notion), developer teams building AI-powered applications (Sourcegraph, Cresta, Quora), and agentic AI builders plus multimodal application developers. Marquee outcomes include Cursor's 13x inference speedup, Notion's latency reduction from 2 seconds to 350 milliseconds, Vercel achieving 90%+ error-free generation at 40x faster throughput, and Factory cutting costs to 6–20% of frontier-model levels. As of May 2026, Fireworks reports processing more than 30 trillion tokens daily for 10,000+ customers at an annualized revenue run-rate of approximately $800 million, with active fundraising discussions at a $15 billion valuation.

Name

Fireworks AI

Legal name

Fireworks AI, Inc.

Website

https://fireworks.ai

Company type

Private

Founded year

2022

Operating status

Operating

Headcount range

101–250 employees

Short description

Fireworks AI operates a high-performance AI inference and training platform for open-source models, founded by the former Meta PyTorch engineering team. It serves AI-native startups, enterprises, and developer teams with serverless and dedicated GPU deployments, processing 30+ trillion tokens daily across 10,000+ customers.

Ownership category

akta.pro rank

Product category

AI Inference Platform

NAICS

Software Publishers (5132), Computer Systems Design and Related Services (54151)

SIC

Services-Prepackaged Software (7372)

akta.pro primary industry

End-to-End Enterprise AI Platforms (MLOps & Model Lifecycle Management) (HDAEANAA)

akta.pro secondary industry

GPU-Accelerated & AI Training/Inference Servers (HDACABAG)

Keywords

• AI inference platform
• Generative AI infrastructure
• LLM deployment
• Model serving platform
• AI cloud platform

Headquarters

HQ city

Redwood City

HQ country

United States

HQ region

North America

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

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

Revenue model

1. Serverless Inference: Pay-per-token pricing for serverless inference. Cached input tokens priced at 50%. Batch inference also at 50% of serverless pricing. Primary revenue driver for developers and startups.
2. On-Demand Deployments: Dedicated GPU deployments with per-GPU-second billing. H100 80GB at $7/hr, H200 141GB at $7/hr, B200 180GB at $10/hr, B300 288GB at $12/hr. No extra charges for startup times.
3. Reserved Capacity: Guaranteed capacity with higher quotas, early access to newest regions and hardware. Supports BYOC deployment flexibility. Targets enterprise customers.
4. Fine-Tuning Services: LoRA and full-parameter training priced per million training tokens or per GPU hour for reinforcement fine tuning. Base models up to 16B: $0.50-2.00 per 1M tokens depending on method.
5. Enterprise Contracts: Direct enterprise sales for reserved capacity and custom deployments. Reported $800M annualized revenue as of May 2026.

Pricing tiers

Go-to-market motion3 records

Distribution channels5 records

Marketing channels7 records

Core offering

Fireworks AI provides a unified AI inference and training platform that runs and fine-tunes 1000+ open-source generative AI models at production scale. The platform offers serverless pay-per-token inference, on-demand dedicated GPU deployments, and reserved enterprise capacity, combined with a Training Platform supporting full-parameter SFT/DPO/RFT. Customers access the service via an OpenAI and Anthropic-compatible API, processing 30+ trillion tokens daily across 14 regions.

Product overview

Fireworks AI is an AI inference and training platform founded by former Meta PyTorch engineers. The platform provides a unified infrastructure for running and fine-tuning open-source AI models, consisting of: (1) Fireworks Inference Platform for serverless, on-demand, and reserved inference with up to 4x throughput gains; (2) Fireworks Training Platform with Training Agent, Managed Training, and Training API for full-spectrum model customization; (3) Model Library with 1000+ open-source models including DeepSeek, Kimi, Qwen, GLM, and NVIDIA Nemotron with day-zero support; (4) Specialized solutions including Code Assistance, Agentic Systems, Multimodal AI, and Enterprise RAG; (5) Acquired Hathora container orchestration platform for real-time workloads. The platform is OpenAI and Anthropic API compatible for easy migration.

Differentiator

Problem solved

Functional benefit

Products and services

• Fireworks Inference Platform High-performance AI inference platform optimized from custom kernels to memory management, delivering up to 4x higher throughput and industry-leading latency. Supports serverless, on-demand, and reserved capacity deployment options for developers and enterprise AI teams running open-source generative models in production.
• Fireworks Training Platform

Quantifiable outcome

• Cursor achieves 13x faster inference speed using Fireworks speculative decoding
• +8 more outcomes

Named customers25 records

Segments5 records

Ideal customer profiles4 records

Technology focussed Yes

API detail

Has API

Yes

API docs

API detail

Core technology

AI maturity

App detail

Integration4 records

AI capability10 records

Feature8 records

Partnerships

Eight partnerships are on record, tiered core and minor.

• MicrosoftcoreTechnology or IntegrationFireworks AI general availability announced in Microsoft Foundry at Build 2026. Enables enterprise deployment through Azure cloud with unified platform experience. Integration with Microsoft IQ context intelligence layer.
• MongoDB for StartupscoreGTM or Marketing PartnerFireworks AI became launch partner in MongoDB for Startups program expansion in January 2026. Program encompasses companies representing $200B+ combined valuation. Provides matched credit offers across complementary technologies.
• NVIDIAcoreTechnology or IntegrationDay-zero availability of NVIDIA Nemotron 3 Super model announced. NVIDIA GTC 2026 featured conversation between Jensen Huang and CEO Lin Qiao. Deep Blackwell platform optimization partnership for 4-10X cost reductions.
• TemporalminorGTM or Marketing PartnerLaunch partner in MongoDB for Startups program alongside Fireworks AI. Provides workflow orchestration complementary to Fireworks inference platform.
• HathoracoreStrategic or Co-development PartnerAcquired by Fireworks AI in March 2026. Hathora built global container orchestration platform for latency-sensitive real-time workloads across 14 regions, multiple bare-metal providers, and four clouds. Team joining to enhance compute orchestration for AI inference.
• CursorcoreTechnology or IntegrationCursor uses Fireworks as primary inference provider for Composer 2 model. 'Way more performant than open source engines and is what we use in production. RL inference scales elastically and globally because of it.' Fireworks powers Fast Apply and Copilot++ models.
• VercelcoreTechnology or IntegrationVercel's v0 model uses Fireworks - 'SOTA changes every day, so you don't want to tie yourself to a single model. Using a fine-tuned RL model with Fireworks, we perform substantially better than Sonnet.' CTO Malte Ubl confirmed partnership.
• UiPathcoreTechnology or IntegrationUiPath powers Autopilot and Delegate with open models via Azure Foundry using Fireworks AI. Achieves quality matching Claude Sonnet 4.6 at fraction of cost for Computer Use tasks.

Scale indicators10 records

Recent moves6 records

Expansion highlights7 records

Direct peers

• Modal Labs: Modal offers serverless cloud infrastructure for AI/ML workloads including LLM inference, with a developer-first API and Python-native deployment model. It targets the same API-first, open-model inference use cases as Fireworks.
• OpenRouter: OpenRouter provides a unified API that routes inference requests across multiple model providers and self-hosted backends including Fireworks itself. It is an adjacent routing layer but increasingly competes with Fireworks by aggregating open-model supply.
• Replicate: Replicate runs a cloud for running open-source AI models via API, with a developer-first, pay-per-use model. It overlaps directly with Fireworks' serverless inference and model library offering for open-weight models.
• Baseten: Baseten provides dedicated and serverless inference infrastructure for ML models, with strong emphasis on production deployment, scaling, and performance optimization. It directly competes with Fireworks' on-demand and dedicated deployment offerings.
• Together AI: Together AI offers an open-model inference and fine-tuning cloud very similar to Fireworks, targeting developers and enterprises with GPU clusters and an API for open-weight LLMs. It is the closest direct head-to-head competitor in the open-model inference category.
• Anyscale: Anyscale provides AI compute and inference infrastructure built on Ray, including serverless LLM endpoints for open models. It competes directly with Fireworks on serving open-weight models at scale for AI developers and enterprises.

Broad incumbents

• Hugging Face: Hugging Face hosts open models, datasets, and inference endpoints with a strong developer community and growing enterprise tier (Inference Endpoints, Spaces, and dedicated deployments). It overlaps with Fireworks on open-model serving and developer workflow, with broader scope into model distribution.
• Microsoft Azure AI Foundry: Azure AI Foundry is Microsoft's hyperscaler-scale platform for hosting and serving foundation models, where Fireworks is itself listed as a partner. It competes with Fireworks on enterprise inference procurement while also serving as a distribution channel.
• Google Vertex AI: Vertex AI is Google Cloud's enterprise AI platform covering training, tuning, and inference for both Google and select third-party models. It is a broad incumbent offering that competes with Fireworks for enterprise inference workloads, particularly via Google Cloud's data and MLOps stack.
• AWS Bedrock: AWS Bedrock is a managed service offering foundation models (including open and proprietary) with enterprise-grade security, compliance, and AWS-native billing. It is a broader incumbent alternative to Fireworks for enterprises already standardized on AWS.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat6 records

Key risks6 records

Key highlights7 records

Customer concentration

• Xx.com/thefireworksai
• LinkedInlinkedin.com/company/fireworks-ai

Revenue estimate

Valuation estimate

Number of profiles

Profiles1 record

Subsidiaries1 record

Funding overview

Funding rounds7 records

Investors22 records

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

M&A1 record

Investments

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

Frequently asked questions about Fireworks AI