Fireworks AI

TL;DR. Fireworks AI provides AI infrastructure. Inference runs a trained model to produce outputs. Fireworks’ serverless inference lets developers run models while Fireworks manages servers. Dedicated graphics processing unit (GPU) deployments reserve GPUs for one customer. Fine-tuning adapts models with task-specific data. Fireworks best fits technical teams with large production workloads.

At a Glance

Fireworks’ latest disclosed investment is a $1.505 billion Series D funding round. Series D is a later round of investment funding. Fireworks announced it in July 2026 at a $17.5 billion valuation.

The company builds AI infrastructure for open models. Open models have publicly available weights. Weights are numbers that control how a model produces outputs. Fireworks positions itself as an inference and training platform for developers and enterprises. Training means building or adjusting a model.

Fireworks identifies seven co-founders. They are CEO Lin Qiao, Benny Chen, Dmytro Dzhulgakov, Pawel Garbacki, Dmytro Ivchenko, James Reed, and Chenyu Zhao. CNBC reports a 2022 founding date. Fireworks says its founding team brought AI experience from PyTorch, Meta, and Google. PyTorch helps developers build AI models. Fireworks describes a founding belief in specialized AI built on open models that businesses can own and customize.

The platform serves models through application programming interfaces (APIs). It supports dedicated deployments and fine-tuning. The reviewed sources do not establish a current employee total.

Fireworks says Atreides Management, Index Ventures, and TCV led its July 2026 Series D. Evantic Capital, Lightspeed, NVIDIA, 20VC, Bessemer Venture Partners, and Menlo Ventures also participated. Its Series C release reported $327 million in total funding. Series C is an earlier round of investment funding. That figure is from before the Series D.

Key links: Website, documentation, GitHub, and blog.

Adoption & Sentiment

User feedback is divided. Developers praise fast inference, model choice, and API integration. Some users report problems with setup, support, and reliability. AWS Marketplace provides one public review sample. This sample does not represent all users.

AWS Marketplace showed 4.2/5 across 43 ratings in October 2026. AWS listed 39 external reviews and four AWS reviews. AWS labels the total as ratings and the two groups as reviews. The AWS Marketplace page also shows recent reviews from G2, a software review website. These reviews praise inference speed, multiple models, and easy integration.

They also cite unclear advanced configuration, documentation gaps, and the time required to select the right model.

Chiri did not verify current G2, Capterra, or Trustpilot ratings. Unavailable ratings do not establish customer satisfaction.

Fireworks’ Series C release provides an earlier customer total. Its customer pages give adoption examples. Fireworks reports these examples. Heidi, a Fireworks customer, reported moving from a test project to production in four weeks. Heidi also reported latency falling from 25 seconds to seven seconds. Latency is the time that a response takes.

This case shows production use. It does not establish platform-wide adoption. Reviewers also report that users need technical knowledge. They report that teams must test models for each workload. These issues may slow adoption for teams seeking a finished, nontechnical product.

How It Makes Money

In July 2026, Fireworks reported more than $1 billion in annualized revenue run rate. Annualized revenue run rate multiplies current revenue to estimate a yearly rate. The company charges customers for inference, GPU deployments, and model training based on usage.

A token is a unit of model input or output. Standard DeepSeek V4.1 Flash costs $0.30 per million input tokens and $1.20 per million output tokens. Cached input is input that the system stores from an earlier request. Cached input costs $0.006 per million tokens. The reviewed pricing page does not confirm hourly GPU prices.

Self-serve customers prepay credits. Fireworks offers automatic credit purchases, called auto-reload, and monthly spend limits. New users receive a $1 starter credit. Usage pauses after the credits run out unless customers add credits or enable auto-reload.

Fireworks directs enterprise buyers to sales for pricing. Throughput measures processing per unit of time. Fireworks says Reserved Throughput guarantees processing throughput under a service-level agreement (SLA). An SLA states a provider’s service commitments. The reviewed source does not specify a standard commitment period.

Fireworks reported more than $1 billion in annualized revenue run rate in July 2026. CNBC reported that annualized revenue grew fivefold year over year. These are reported annualized revenue run rate figures. They are not audited annual revenue. The reviewed public materials do not estimate the total possible revenue in its market.

Security & Compliance

On October 6, 2026, Fireworks identified a security incident. The incident involved unauthorized use of internal credentials. Credentials are codes that prove identity or grant access. Fireworks says a third party accessed environment variables in evaluation jobs across a limited number of customer accounts. Environment variables are configuration values that a program can read. Evaluation jobs are tasks that test model outputs.

Fireworks believes the third party captured some stored API keys and access tokens. API keys and access tokens are codes that grant access to services. It reports no detected evidence of access to customer models, datasets, training data, or inference traffic. It also reports no detected evidence of access to personal information within its security perimeter. Fireworks says its investigation continues.

Fireworks says it revoked identified credentials. It says it notified affected customers.

Fireworks reports SOC 2 Type II certification. SOC 2 Type II is an audit of security controls over a period of time. It also reports HIPAA compliance. HIPAA is a US law that protects health information. It reports ISO 27001, ISO 27701, and ISO 42001 certifications. These standards cover information security, privacy management, and AI management, respectively.

Fireworks says its controls align with the requirements of GDPR and CCPA. GDPR means General Data Protection Regulation, a European privacy law. CCPA means California Consumer Privacy Act, a California privacy law.

Fireworks says it encrypts data during transmission with Transport Layer Security (TLS) 1.2 or higher. Fireworks says it encrypts stored data with Advanced Encryption Standard with 256-bit keys (AES-256). Encryption converts readable data into protected data that needs a key to read. Fireworks supports customer-managed encryption keys for some training data.

LoRA means low-rank adaptation, a method that trains a smaller set of model weights. Supervised fine-tuning (SFT) trains models with examples of desired outputs. Direct preference optimization (DPO) trains models with preferred outputs. Reinforcement fine-tuning (RFT) trains models using task feedback.

Fireworks limits these keys to datasets and saved training states (checkpoints) for LoRA SFT, DPO, and RFT. This support excludes training that changes all model weights (full-parameter training). Fireworks does not yet support these keys for final model weights.

Fireworks says customers can use their own cloud buckets for certain training workflows. A cloud bucket is an online storage container.

Fireworks says open-model inference has zero data retention by default. This means it does not keep prompts or outputs after processing. Fireworks says it keeps request metadata that it needs to operate and bill the service. Request metadata describes a request, such as its time or token count.

The Responses API is a Fireworks API that manages conversations with multiple exchanges. Fireworks says this API stores conversations for 30 days when store=True, its default. Fireworks says customers can set store=False or delete records. It says enterprise administrators can block features that store customer content. Fireworks says it does not train its models on customer inputs unless the customer explicitly agrees.

Fireworks says enterprise accounts include single sign-on (SSO) and audit logs. SSO lets users access multiple services with one login. Audit logs record system actions. Fireworks says these logs cover storage reads, writes, and deletes.

Fireworks documents access controls across its environments. Access controls limit which actions a user or service can perform.

Fireworks says enterprise administrators can restrict account-wide inference to a selected region. The documentation lists only the US option. Fireworks says the platform rejects requests for processing outside the selected region.

Bring-your-own-cloud (BYOC) runs deployments in the customer’s cloud account. Nexus Auto is a Fireworks feature that can send requests to third-party model providers. Fireworks says a residency setting blocks training, Nexus Auto, and BYOC requests. This setting does not move existing deployments. Existing deployments outside the region can continue to run and incur charges.

Fireworks says its Trust Center provides audit reports and certification documents. The Trust Center is its security information page.

Leadership & Team

Fireworks identifies seven co-founders. CEO Lin Qiao and President George Hu lead the company.

CNBC reported about 200 employees in July 2026. Qiao told CNBC that the company planned to reach 600 by year-end. That figure is a target, not a reported number of employees. Fireworks describes its founding belief as specialized AI built on open models and owned by customers. That is a statement about its product. It does not describe its internal culture in detail.

Products & Features

Fireworks combines hosted model inference, dedicated deployments, and fine-tuning on one platform. Its catalog lists more than 100 models across text, vision, audio, image generation, and embeddings. Embeddings are numerical representations of content that systems use for search and comparison.

Serverless inference runs models on servers that Fireworks shares among customers. Users do not set up GPUs. Standard mode targets default usage. Priority mode targets reliability during busy periods. Fast mode targets requests that need low latency.

For example, DeepSeek V4.1 Flash Standard costs $0.30 per million input tokens. Cached input costs $0.006 per million tokens. Output costs $1.20 per million tokens.

Batch inference processes groups of requests together. It costs 50% of Standard mode prices for input and output tokens.

Rates vary by model and mode.

On-demand deployments are dedicated GPU deployments. They run models on dedicated GPUs. Users choose the GPU type and count for each deployment. They control automatic scaling, which adjusts the number of running model copies. A model copy is one running instance of a model.

Fireworks bills on-demand deployments by GPU-second while model copies run. A GPU-second is one second of GPU use. Deployments support uploaded custom models and trained LoRA add-ons. LoRA means low-rank adaptation, a method that trains a smaller set of model weights.

Fireworks also offers Reserved Throughput through enterprise arrangements. It also offers BYOC options.

Training includes supervised fine-tuning (SFT). SFT trains models with examples of desired outputs. Direct preference optimization (DPO) trains models with preferred outputs. Reinforcement learning (RL) trains models through feedback on their actions. RFT applies RL to fine-tuning.

Fireworks says Managed Training runs standard SFT and DPO jobs. Fireworks lists Managed RFT as paused. The service accepts no new jobs. RL runs only on the Training API.

Python is a programming language. The Training API lets teams write Python training loops. These loops contain code that repeats training steps.

The catalog determines which training methods each model supports. Training and serving use the same platform.

Representational state transfer (REST) is a style for designing APIs. A software development kit (SDK) provides code for calling APIs. A command-line tool accepts typed commands. Developers can use the Fireworks interface, the firectl command-line tool, a REST API, or a Python SDK.

An endpoint is an API address that accepts requests. Inference supports OpenAI Chat Completions and Responses endpoints, plus Anthropic Messages compatibility. Fireworks documentation also lists the OpenAI and Anthropic SDKs, Claude Code, and Codex as compatible clients. A client is software that calls a service. Nexus CLI is Fireworks' command-line interface (CLI) for coding tools.

Fireworks announced on-demand support for NVIDIA Nemotron 3 Ultra on the model’s release day. An inference engine is software that runs model calculations. Fireworks says its FireAttention V4 inference engine is available on B200 GPUs for enterprise customers. B200 is an NVIDIA GPU model.

Classification

Fireworks documents more than 100 supported models. It offers inference, dedicated GPU deployments, and fine-tuning. Chiri places it in AI Infrastructure.

Primary Category: AI Infrastructure Categories: AI Infrastructure, AI Coding Tags: open-source models, LLM inference, model hosting, serverless API, fine-tuning, GPU deployments, multimodal AI, embeddings, OpenAI-compatible API, Anthropic-compatible API, developer tools

Market Impact

Fireworks reported over $1 billion in annualized revenue run rate in July 2026. Fireworks also reported 40 trillion daily tokens. Chiri identifies competition among similar services as a risk. CNBC reported past dependence on a few customers.

Its Series C release reported 10 trillion daily tokens. It also reported 10,000+ customers.

By July 2026, it reported over $1 billion in annualized revenue run rate. This figure was five times the figure for the prior year. Fireworks said more than 95% of its token volume came from models that customers adapted with their own data. Cursor’s coding models and Harvey’s legal AI are examples of this customer-specific use. Fireworks and CNBC reported these operating figures. Annualized revenue run rate is not audited annual revenue.

Together AI is a close open-model competitor. It announced a $305 million Series B in February 2025. Together AI reported 450,000 developers. Together AI reported more than 200 models.

Baseten also competes in inference. It offers cloud, self-hosted, and hybrid deployment options. Self-hosted means the customer runs the software. Hybrid combines cloud and customer-run infrastructure. AWS, Microsoft, and Google offer alternatives within their established cloud services.

Fireworks’ strengths are scale, model customization, and its combined tuning, evaluation, and serving workflow. Token volume from customized models suggests large production workloads. It does not show that these strengths will continue. It does not show that customers find it difficult to change providers.

Risks include increasing computing needs and dependence on a few customers. Fireworks announced plans to expand its computing resources globally after its Series C funding. CNBC described Cursor’s 2025 revenue share as “about half”. CEO Lin Qiao said the customer mix became more diverse after that. CNBC did not quantify current dependence on individual customers.

Similar services compete on price. Open-source inference software gives customers more alternatives. Open-source means software with public source code. Cloud providers also include model hosting within their services.

Fireworks best fits technical teams with large production workloads. These teams use proprietary data and need to customize open models.

Teams should compare it with Together AI and Baseten. Teams should also compare cloud platforms when they want to buy all services through one contract. Chiri recommends that teams consider Fireworks for production workloads. Teams should then validate latency, total cost, support, and computing resources for the specific use case.

Chiri sees possible growth. This possible growth depends on more types of customers using Fireworks. It also depends on Fireworks keeping its technology different from competitors.

Who It's Built For

Fireworks serves startups that build AI products. It also serves enterprises that scale AI across products, teams, and customers. Its clearest fit is teams that need custom models and inference capacity that can increase with demand. These teams also need control over cost or data. The reviewed sources do not give customer counts by size or region.

Startup founders and AI product builders can begin with serverless models. They can add fine-tuning and Reserved Throughput as usage grows. Enterprise buyers include AI and platform engineers, developers, and security teams. Fireworks also targets product and innovation leaders. These leaders want to put AI into workflows. They do not want to build serving infrastructure.

Common use cases include coding agents and code completion. Code completion suggests code as a developer types. Coding agents use models to perform programming tasks. Other uses include conversational assistants and workflows that have multiple steps. Semantic search finds results by meaning. Enterprise knowledge retrieval finds information in company documents.

Enterprise knowledge retrieval often uses retrieval-augmented generation (RAG). RAG bases model answers on a company’s documents.

Customer examples show these uses in software development (Cursor and Vercel), workplace software (Notion), and contact centers (Cresta). They also show clinical documentation (Heidi Health). Fireworks also presents legal AI through Harvey and research agents through Genspark.

These examples show use in several industries. They do not establish a measured ranking of adoption.

Customers range from early-stage startups to large enterprises. Notion’s Fireworks case study describes AI workflows. According to the case study, these workflows serve more than 100 million users. According to that case study, Notion's customers include nearly 70% of Fortune 100 companies. Fireworks says enterprise deployments are available in more than 30 regions. Fireworks describes global deployment options.

The reviewed sources do not identify a dominant geographic market.

Sources

• Funding announcement

• Company background

• Founding team

• Index Ventures profile

• Marketplace reviews

• Customer examples

• Heidi Health case study

• Security incident bulletin

• Security controls

• Data handling

• Enterprise account controls

• Audit logs

• Data residency

• Serverless pricing

• Billing and credits

• Starter credits

• Dedicated deployments

• Fine-tuning documentation

• API compatibility

• NVIDIA model support

• FireAttention announcement

• Enterprise uses

• Notion case study

• CNBC company profile

• Together AI funding

• Baseten deployments

• Series C funding

Chiri Analysis

Chiri Score: 74/100

Verdict

Best for:

• AI product teams running high-volume inference on open models

• Teams that fine-tune open models on proprietary data and serve them on one platform

• Startups that want to start on serverless inference and scale to Reserved Throughput

• Developers migrating from OpenAI or Anthropic APIs who want compatible endpoints

• Coding, agent, and RAG workloads, and workloads that combine text, images, or audio

Not for:

• Nontechnical teams seeking a finished AI product that does not require programming

• Buyers who require independently verified compliance certifications before evaluation

• Security-sensitive teams that will not wait for the end of the October 2026 incident investigation

• Organizations that require one contract with an existing large cloud provider

• Teams that need fixed, published enterprise pricing without contacting sales

Head-to-head

Frequently Asked Questions

Is Fireworks AI SOC 2 compliant?

Fireworks reports SOC 2 Type II certification. Fireworks also reports HIPAA compliance and ISO 27001, ISO 27701, and ISO 42001 certifications. Chiri Atlas did not independently verify these claims. Buyers should request audit reports through the Fireworks Trust Center.

How much does Fireworks AI cost?

Fireworks uses usage-based pricing. Self-serve customers prepay credits.

Serverless inference bills per token, a unit of model input or output. Standard DeepSeek V4.1 Flash costs $0.30 per million input tokens and $1.20 per million output tokens. Cached input costs $0.006 per million tokens. Batch inference costs 50% of Standard mode rates for input and output. Fireworks directs enterprise buyers to sales for pricing.

Chiri did not verify hourly GPU prices.

Did Fireworks AI have a security incident?

Yes. On October 6, 2026, Fireworks identified unauthorized use of internal credentials. Fireworks says a third party accessed environment variables in evaluation jobs in a limited number of customer accounts. Fireworks believes the third party captured some stored API keys and access tokens. API keys and access tokens are codes that grant access to services.

It reports no detected evidence of access to customer models, datasets, training data, or inference traffic. It also reports no detected evidence of access to personal information within its security perimeter. Fireworks says it revoked identified credentials. Fireworks says it notified affected customers. Fireworks says its investigation continues.

Who are Fireworks AI's competitors?

Fireworks competes with Together AI and Baseten in open-model inference. AWS, Microsoft, and Google offer bundled alternatives through their cloud platforms. The report says Together AI offers more than 200 models. Fireworks documents more than 100 models. The report says Baseten offers cloud, self-hosted, and hybrid deployment options.

Is Fireworks AI good for enterprise?

Fireworks markets an enterprise offering with Reserved Throughput. Enterprise buyers contact sales for pricing. Fireworks says enterprise accounts include single sign-on (SSO) and audit logs. Fireworks reports its compliance certifications. Chiri Atlas did not verify them. The October 2026 credential incident remains under investigation.

Chiri recommends that enterprises complete a full security review before they sign a contract with Fireworks.

Does Fireworks AI support fine-tuning?

Yes. Fireworks offers supervised fine-tuning (SFT) and direct preference optimization (DPO) through Managed Training. Managed RFT (reinforcement fine-tuning) accepts no new jobs. Reinforcement learning (RL) runs only on the Training API. The catalog determines which training methods each model supports.

Is the Fireworks AI API compatible with OpenAI and Anthropic?

Yes. Fireworks says it supports OpenAI-compatible and Anthropic-compatible API endpoints. Its documentation covers integrations with the OpenAI and Anthropic software development kits (SDKs). Chiri expects this compatibility to reduce the work needed to move existing applications to Fireworks.

Who founded Fireworks AI?

Fireworks identifies seven co-founders. They are Lin Qiao, Benny Chen, Chenyu Zhao, Dmytro Dzhulgakov, Dmytro Ivchenko, James Reed, and Pawel Garbacki. Lin Qiao is CEO and previously led PyTorch at Meta. CNBC reports a 2022 founding date.

---

Reviewed by Chiri Atlas Research Desk (AI Tooling Analyst) on 2026-10-25.