Fireworks AI — research brief

Slug: fireworks-ai
Date: 2026-09-21
Industry: AI Inference / Specialized Intelligence Platform
Primary sources: fireworks.ai pricing, docs.fireworks.ai, fireworks.ai/inference, company Series D blog (2026-07-15)

Labels

Company snapshot

Funding & scale (company-stated)

Named customers (company testimonials / blog)

Cursor, Harvey, Notion, Vercel, Quora, Sourcegraph, Cresta, UiPath, Genspark, Gumloop, Motif, rLLM (FACT — fireworks.ai homepage / inference testimonials + Series D blog names Cursor & Harvey).

Product & pricing teaching arc (FACT — pricing + docs)

  1. Serverless Inference (Standard) — pay per token; $1 free credits on pricing page; OpenAI/Anthropic compatible; no cold starts (docs).
  2. Cached input — default ~50% of input on text/vision unless model lists otherwise (docs).
  3. Batch — 50% of Standard serverless on input+output (docs/pricing).
  4. Priority — service_tier: "priority" premium reliability; e.g. GLM 5.2 Standard $1.40/$0.14/$4.40 → Priority $1.75/$0.18/$5.50 per 1M (docs serverless pricing).
  5. Fast — Fast model IDs for higher speed; e.g. GLM 5.2 Fast $2.10/$0.21/$6.60 (docs).
  6. On-Demand Deployments — pay per GPU-second/hour: H100/H200 $8.00/hr; B200 $13.00; B300 $15.00; GB300 $20.00 (fireworks.ai/pricing). Region-restricted 1.5× (pricing note).
  7. Training — Managed LoRA/full SFT & DPO per 1M training tokens by model size bands (e.g. ≤16B LoRA SFT $0.50/1M); RFT billed per GPU-hour at on-demand rates; Serverless Training API token meters; Dedicated Training API GPU-hours (pricing).
  8. Reserved Capacity / Enterprise — guaranteed capacity, higher quotas, BYOC flexibility; Contact sales (inference page). Reserved Throughput adds SLA-backed throughput on serverless (docs).

Example serverless Standard rates (docs, $/1M in/cached/out)

Security claims (FACT — inference FAQ)

SOC 2 Type II, HIPAA-ready, GDPR-compliant; does not use inference I/O for other purposes (company FAQ).

What we do NOT invent

Private gross margin, headcount, utilization %, unpublished enterprise contract rates, or ARR beyond the company blog claim.

Accent

#082dff — official mark fill from Fireworks logo SVG.

# Fireworks AI — research brief

**Slug:** fireworks-ai  
**Date:** 2026-09-21  
**Industry:** AI Inference / Specialized Intelligence Platform  
**Primary sources:** fireworks.ai pricing, docs.fireworks.ai, fireworks.ai/inference, company Series D blog (2026-07-15)

## Labels
- **FACT** — from Fireworks pricing, docs, or official blog
- **SECONDARY** — press repeating company claims
- **INFERENCE** — structural read from public product shape (not a private P&L)

## Company snapshot
- Platform for training and serving specialized / open models (FACT — fireworks.ai homepage + inference page).
- Co-founder & CEO: Lin Qiao (PyTorch Meta background commonly cited in SECONDARY press).
- Positioning: “specialized intelligence” — train on your data, serve on Fireworks stack (FACT — Series D blog).

## Funding & scale (company-stated)
- **Series D:** $1.505B at $17.5B valuation, announced ~2026-07-15 (FACT — fireworks.ai/blog/series-d-announcement).
- Led by Atreides Management, Index Ventures, TCV; participants include Nvidia, Lightspeed, Evantic, Bessemer, Menlo, 20VC, others (FACT — company blog / BusinessWire narrative).
- Company claims **>$1B annualized revenue run rate** and **>40 trillion tokens/day**, with **>95%** of tokens from models specialized on customer data (FACT claim — Series D blog). Do not invent other ARR/headcount.

## Named customers (company testimonials / blog)
Cursor, Harvey, Notion, Vercel, Quora, Sourcegraph, Cresta, UiPath, Genspark, Gumloop, Motif, rLLM (FACT — fireworks.ai homepage / inference testimonials + Series D blog names Cursor & Harvey).

## Product & pricing teaching arc (FACT — pricing + docs)
1. **Serverless Inference (Standard)** — pay per token; $1 free credits on pricing page; OpenAI/Anthropic compatible; no cold starts (docs).
2. **Cached input** — default ~50% of input on text/vision unless model lists otherwise (docs).
3. **Batch** — 50% of Standard serverless on input+output (docs/pricing).
4. **Priority** — `service_tier: "priority"` premium reliability; e.g. GLM 5.2 Standard $1.40/$0.14/$4.40 → Priority $1.75/$0.18/$5.50 per 1M (docs serverless pricing).
5. **Fast** — Fast model IDs for higher speed; e.g. GLM 5.2 Fast $2.10/$0.21/$6.60 (docs).
6. **On-Demand Deployments** — pay per GPU-second/hour: H100/H200 $8.00/hr; B200 $13.00; B300 $15.00; GB300 $20.00 (fireworks.ai/pricing). Region-restricted 1.5× (pricing note).
7. **Training** — Managed LoRA/full SFT & DPO per 1M training tokens by model size bands (e.g. ≤16B LoRA SFT $0.50/1M); RFT billed per GPU-hour at on-demand rates; Serverless Training API token meters; Dedicated Training API GPU-hours (pricing).
8. **Reserved Capacity / Enterprise** — guaranteed capacity, higher quotas, BYOC flexibility; Contact sales (inference page). Reserved Throughput adds SLA-backed throughput on serverless (docs).

### Example serverless Standard rates (docs, $/1M in/cached/out)
- GLM 5.2: $1.40 / $0.14 / $4.40
- Kimi K2.7 Code: $0.95 / $0.19 / $4.00
- MiniMax M3: $0.30 / $0.06 / $1.20
- OpenAI GPT OSS 120B: $0.15 / $0.015 / $0.60
- Embeddings: $0.008–$0.016 / 1M (size bands); Qwen3 8B $0.10

## Security claims (FACT — inference FAQ)
SOC 2 Type II, HIPAA-ready, GDPR-compliant; does not use inference I/O for other purposes (company FAQ).

## What we do NOT invent
Private gross margin, headcount, utilization %, unpublished enterprise contract rates, or ARR beyond the company blog claim.

## Accent
`#082dff` — official mark fill from Fireworks logo SVG.