Open models built for your AI agents
Fireworks is the inference and training platform for AI agents. Genspark’s research agent made 33% more tool calls than a state-of-the-art closed model, with 12% better quality, at 50% lower cost.
Fireworks is the inference and training platform for AI agents. Genspark’s research agent made 33% more tool calls than a state-of-the-art closed model, with 12% better quality, at 50% lower cost.
Teams running agents on Fireworks
Start with inference on open models. Add training later if your tool-call patterns need it.
Start on serverless per token, then move to on-demand or Reserved Throughput as load grows; Sentient served 5.6M+ queries in a week with zero degradation.
If an off-the-shelf model misses your tool-call patterns, fine-tune on multi-turn tool calls with SFT or RFT; Genspark went from POC to production in 4 weeks with RFT on Kimi K2.
Orchestrate across models through the agent frameworks that connect to Fireworks.
Why use Fireworks for AI agents?
OpenAI-compatible tool calls, plus parallel calls on models that support them.
Connect MCP tools and cap how many tool calls one response can make.
Trilogy runs at ~1/5 the cost of proprietary systems, with 93.8% cache hits.
Notion cut latency from about 2s to 350ms for 100+ million users.
Pick from the recommended models for agents with tool use, such as GLM 5.2 or Kimi K2.6.
Describe each function in the OpenAI-compatible tools format and set whether and which tools the model can call.
Turn on JSON mode or Grammar mode so every step returns output your code can parse.
Plug Fireworks into a supported agent framework, or call MCP tools directly through the Responses API.
On a dedicated deployment, autoscale on tokens/s, requests/s or concurrent requests, and scale to zero when idle.
Your agent sends prompts and tool results on every call, and you decide where that data sits and who holds the keys.
Fireworks keeps no request data by default. If you train with Secure RFT, your reward functions stay with you.
Fireworks covers SOC 2 Type II and HIPAA, plus ISO 27001, 27701 and 42001.
Fireworks encrypts data with AES-256 and TLS 1.2+. You can bring your own storage bucket or encryption key.
Enterprise accounts get audit and access logs for compliance reviews. Other account tiers don’t include them.
Set a p95 latency budget for each call, then test every model you shortlist against it. If your budget depends on speculative decoding, plan for a dedicated deployment, the only place it runs. In a voice-to-action workflow, RADPAIR cut report turnaround from 15–20s to 2–5s.
A failed tool call means a retry, and every retry adds to what the task costs. Reliability shows up on the bill. Use JSON or Grammar mode, and give each step only the tools it needs.
Add up spend per completed task across every call. Innovative Solutions runs 4–10B tokens a month and moved 90% of its model spend to Fireworks in about 1.5 weeks. A full-parameter fine-tune runs on a dedicated deployment billed per GPU-second, so budget it apart from per-token traffic.
Yes. Function calling and tool calling are the same thing, and Fireworks runs both through its OpenAI-compatible API. Parallel calls depend on the model you pick.
Start with a current serverless model with tool calling, such as Kimi K3 or GLM 5.3. Each model card lists its context length and whether it supports parallel tool calls.
The agent framework integrations docs list every supported framework. If you skip a framework, step 4 above covers the Responses API route.
Yes. JSON mode holds output to a schema you supply, and Grammar mode holds it to a grammar you define.
On serverless, you pay per 1M tokens, with separate input, cached input and output rates for each model. Batch runs at 50% of serverless.
Not by default, and audit logs are Enterprise-only. The Security section above covers retention, encryption, keys and certifications.
Run your agents on open models built for tool calling, then plan your deployment with our team.
