Fireworks AI Review 2026 — Fastest Open-Source LLM Inference API

Features

• Run open models with serverless inference and no GPU setup
• Deploy models on dedicated GPUs with per-second billing
• Fine-tune models with LoRA, DPO, SFT, and reinforcement fine-tuning
• Serve hundreds of fine-tuned variants using Multi-LoRA
• Build agents with JSON mode, grammar mode, and tool calling
• Use text, vision, audio, speech, image, and video models
• Access OpenAI- and Anthropic-compatible inference APIs

Use cases

• Build coding assistants and code-generation products
• Run conversational AI and customer-support systems
• Create agents that call APIs and retain context
• Add semantic search and enterprise RAG to applications
• Fine-tune models on proprietary business data
• Process images, documents, audio, and video

Pros

Cons

Latest updates

• 2026-11-04 Cookbook migration and first PyPI release (fireworks-training-cookbook 2026.11.5.1)

  The public cookbook now lives at fw-ai-external/cookbook; fireworks-training-cookbook 2026.11.5.1 is available on PyPI.

• 2026-11-03 New Serverless mode: Ultra-fast (Preview)

  Ultra-fast is a new high-speed Serverless mode, now in Preview, for highly latency-sensitive applications.

• 2026-11-02 Serverless Training deprecation: DeepSeek V4 Flash (0731)

  DeepSeek V4 Flash (0731) is removed from the shared Serverless Training pool effective November 2, 2026.

• 2026-10-27 New model on Serverless: Rune 26B-A4B

  Rune 26B-A4B is now available on Serverless.

• 2026-10-21 Training SDK 1.2.30: clearer TITO errors and trajectory discard (1.2.30)

  Errors identify their source; unparseable trajectories are discarded; failure traces are saved locally; sample supports image inputs.

Capabilities

• Command line — “API, SDK, CLI, console” source
• Choice of models — “Intelligently orchestrate models for work and applications.” source
• Self-hosted — “Yes. Fireworks Sovereign brings our serving stack to enterprise infrastructure in your own cloud account or data center.” source
• No code retained — “Retention depends on the features you use, with controls to manage stored data and enforce zero data retention for supported workloads.” source
• API — “API, SDK, CLI, console” source
• Official SDKs — “API, SDK, CLI, console” source
• Runs models for you — “Serve open and custom models” source
• Traces and evaluates — “For core workloads that encode your intellectual property, evaluate models against your own benchmarks and train specialized models to outperform general-purpose alternatives on the tasks that matter to your product.” source

Get it

• Documentation

Security

• SOC 2 Type II — “SOC 2 Type 2” source
• SOC 2 — “SOC 2 Type 2” source
• ISO 27001 — “ISO 27001 Certificate” source
• ISO 42001 — “ISO 42001 Certificate” source
• GDPR — “Compliance SOC 2 Type 2 HIPAA GDPR” source
• HIPAA — “SOC 2 Type 2 HIPAA” source

Pricing

Prices checked

2026-10-23

Serverless Inference — Embeddings

• $0.008 per 1M input tokens (up to 150M parameters)
• $0.016 per 1M input tokens (150M - 350M)
• $0.1 (Qwen3 8B)

• Usage-based billing on input, output, and cached tokens
• Pre-paid credits; usage is deducted from your balance

Managed Training

• $0.50 per 1M training tokens (models up to 16B parameters)
• $1.00 per 1M training tokens (models up to 16B parameters)
• $2.00 per 1M training tokens (models up to 16B parameters)
• $3.00 per 1M training tokens (models 16.1B - 80B)
• $6.00 per 1M training tokens (models 16.1B - 80B)
• $12.00 per 1M training tokens (models 16.1B - 80B)
• $6.00 per 1M training tokens (models 80B - 300B)
• $12.00 per 1M training tokens (models 80B - 300B)
• $24.00 per 1M training tokens (models 80B - 300B)
• $10.00 per 1M training tokens (models >300B)
• $20.00 per 1M training tokens (models >300B)
• $40.00 per 1M training tokens (models >300B)

• Supervised and preference fine-tuning
• Prices shown per 1M training tokens

Serverless Training API

• $6.19 per 1M Prefill tokens — Muse Spark Open
• $1.239 per 1M Cached Prefill tokens — Muse Spark Open
• $15.48 per 1M Sample tokens — Muse Spark Open
• $18.58 per 1M Train tokens — Muse Spark Open
• $0.60 per 1M Prefill tokens — DeepSeek V4.1 Flash
• $0.12 per 1M Cached Prefill tokens — DeepSeek V4.1 Flash
• $3.60 per 1M Sample tokens — DeepSeek V4.1 Flash
• $4.98 per 1M Train tokens — DeepSeek V4.1 Flash
• $2.96 per 1M Prefill tokens — GLM 5.3 Flash
• $0.593 per 1M Cached Prefill tokens — GLM 5.3 Flash
• $7.41 per 1M Sample tokens — GLM 5.3 Flash
• $8.89 per 1M Train tokens — GLM 5.3 Flash
• $4.86 per 1M Prefill tokens — GLM 5.3
• $0.972 per 1M Cached Prefill tokens — GLM 5.3
• $12.15 per 1M Sample tokens — GLM 5.3
• $14.58 per 1M Train tokens — GLM 5.3
• $1.86 per 1M Prefill tokens — Qwen 3.8 27B
• $0.372 per 1M Cached Prefill tokens — Qwen 3.8 27B
• $5.595 per 1M Sample tokens — Qwen 3.8 27B
• $4.103 per 1M Train tokens — Qwen 3.8 27B
• $10.87 per 1M Prefill tokens — Kimi K3
• $2.17 per 1M Cached Prefill tokens — Kimi K3
• $27.11 per 1M Sample tokens — Kimi K3
• $32.55 per 1M Train tokens — Kimi K3
• $1.74 per 1M Prefill tokens — DeepSeek V4 Flash 0731
• $0.35 per 1M Cached Prefill tokens — DeepSeek V4 Flash 0731
• $4.33 per 1M Sample tokens — DeepSeek V4 Flash 0731
• $5.20 per 1M Train tokens — DeepSeek V4 Flash 0731
• $1.96 per 1M Prefill tokens — Muse Glimmer 30B
• $0.39 per 1M Cached Prefill tokens — Muse Glimmer 30B
• $4.88 per 1M Sample tokens — Muse Glimmer 30B
• $5.86 per 1M Train tokens — Muse Glimmer 30B

• LoRA training on launch models
• Pay for tokens prefetched, sampled, and trained
• No provisioning or idle cost

On-Demand Pricing

• $0.134 per minute — H100 80 GB GPU
• $8.00 per hour — H100 80 GB GPU
• $0.134 per minute — H200 141 GB GPU
• $8.00 per hour — H200 141 GB GPU
• $0.217 per minute — B200 180 GB GPU
• $13.00 per hour — B200 180 GB GPU
• $0.250 per minute — B300 288 GB GPU
• $15.00 per hour — B300 288 GB GPU
• $0.334 per minute — GB300 288 GB GPU
• $20.00 per hour — GB300 288 GB GPU

• On-demand deployments
• No extra charges for start-up times