Fireworks AI
Free tierThe fastest inference platform for generative AI — train, tune, and deploy open models at production scale.
Free tier available·All audiences·Powered by Multiple open-source models (DeepSeek, Llama, Mistral, Qwen, GLM, Kimi, Gemma, etc.) plus closed-model API compatibility (OpenAI, Anthropic)·API available
Key strengths
High-throughput, low-latency inference optimized at every layerFine-tuning and RL training on open modelsOpenAI and Anthropic API compatibility (drop-in replacement)Serverless and on-demand GPU deploymentsMulti-LoRA and post-trained model supportMulti-region deployments
Free tier + paid plans · from $0.07 USD/mo
San Francisco, USA
Founded 2022
Self-hostable
Facts verified against fireworks.ai on
Recent changes
- Models updated: Multiple open-source models (DeepSeek, Llama, Mistral, Qwen, GLM, Kimi, Gemma, etc.) plus closed-model API compatibility (OpenAI, Anthropic)
- Plans updated: Serverless inference billed per token (input/output/cached); embeddings from $0.008/1M tokens; training from $0.50/1M tokens (LoRA SFT, models up to 16B); on-demand GPU from $8.00/hr (H100) to $20.00/hr (GB300).
- Starting price: $0.07/mo → $0.008/mo
Get alerts when Fireworks AI changes
Free. We'll email you when pricing, the free tier, API access or open-source status changes.
No ratings yet
Fireworks AI — Technical Use Cases
Fireworks AI is purpose-built for AI developers and engineering teams targeting the following workloads:
- Code assistance: Low-latency inference on code-specialized models (e.g., DeepSeek, Qwen) integrated with tools like Cursor and Sourcegraph.
- Conversational AI & agentic systems: High-throughput serverless inference for multi-turn dialogue and autonomous agent pipelines requiring fast, reliable token generation.
- Enterprise RAG: Embeddings (from $0.008/1M tokens) combined with serverless inference for retrieval-augmented generation pipelines at scale.
- Multimodal applications: Access to multimodal model families within the same inference API surface.
- Fine-tuned / specialized models: LoRA SFT training on domain-specific data (models up to 16B parameters) with multi-LoRA serving — run multiple adapters from one base model deployment.
- Search & semantic retrieval: Embedding generation at scale for vector search applications.
- Drop-in OpenAI/Anthropic replacement: Migrate existing applications to open-source models without rewriting API client code.
- On-demand GPU workloads: Burst or sustained compute on H100 ($8.00/hr) or GB300 ($20.00/hr) instances for training or batch inference jobs.
Alternatives
See all alternatives Alternatives to Fireworks AI
Is Fireworks AI yours? Claim the listing & get the badge