Fireworks AI Inference Platform

Offering overview

Fast serving is the core promise: open-weight models such as gpt-oss, Kimi K2 and K3, DeepSeek V3 and V4 Flash, GLM 5.2 and 5.3, Qwen3, MiniMax M3 and Gemma 4 are exposed through OpenAI- and Anthropic-compatible APIs with priority and fast tiers, and the in-house Ember-1 model is also available. Teams that need isolation can spin up on-demand deployments on H100, H200, B200, B300 or GB300 hardware, or reserve capacity for guaranteed quotas. A Training API offers guided, configuration-led or fully custom training loops, with LoRA and full fine-tuning billed per million training tokens. Deployments can be pinned to 26 single regions spanning the United States, Canada, Germany, Iceland, Japan, Malaysia and Australia, or to multi-region pools, though region-restricted deployments cost 1.5x. Product teams building latency-sensitive assistants, coding agents and voice applications are the typical buyers. Published rates begin at $0.008 per million input tokens for small embedding models; on-demand H100 and H200 GPUs cost $8.00 per hour and LoRA fine-tuning of models up to 16B starts at $0.50 per million training tokens.

Technical specifications

Containers

Unknown

Kubernetes

Unknown

Root access

Unknown

API access

Unknown

Autoscaling

Unknown

Persistent processes

Unknown

AI capabilities

Model serving

gpt-oss-120B, Kimi K3, Kimi K2, DeepSeek V4 Flash, DeepSeek V3, GLM 5.3, GLM 5.2, Qwen3, MiniMax M3, Gemma 4, Ember-1

AI frameworks

Not stated

Inference endpoints

Unknown

Fine-tuning

Unknown

Vector database

Unknown

Model catalog

Unknown

Monitoring

Not stated

Evaluation

Not stated

Security & compliance

Geography

Provider headquarters are separate from offering infrastructure and availability.

Available plans

Pricing can change. Promotional prices are shown separately from standard prices.