fireworks-ai-inference

Fireworks AI -- Fast Inference & Fine-Tuning

Fastest open-model inference platform with serverless and on-demand GPU deployments, OpenAI-compatible API, and built-in fine-tuning (SFT, DPO, RL).

When to Use Fireworks AI

Use Fireworks AI when:

• Need fast serverless inference for open-source models (Llama, Qwen, DeepSeek, Mixtral)
• Want OpenAI SDK drop-in replacement with open models
• Need fine-tuning without managing infrastructure (SFT, DPO, RL)
• Require structured output / JSON mode / function calling with open models
• Need dedicated GPU deployments with predictable latency
• Require SOC2 or HIPAA compliance
• Want prompt caching and batch inference for cost savings

Use alternatives instead:

Need	Use Instead
Self-hosted inference (full control)	vLLM, TensorRT-LLM
Cheapest serverless inference	Groq (free tier), Together AI
Managed LoRA fine-tuning (no infra)	Tinker
Closed-model APIs (GPT-4, Claude)	OpenAI, Anthropic direct
GPU instances with SSH access	Lambda Labs, RunPod
Multi-cloud orchestration	SkyPilot

Credential Setup

Credentials are auto-injected by openscience when connected via the dashboard.

# Verify credentials
[ -n "$FIREWORKS_API_KEY" ] && echo "FIREWORKS_API_KEY set" || echo "NOT SET"

If not set: connect Fireworks AI at https://app.syntheticsciences.ai -> Services, then restart openscience.

Quick Start

Install

pip install fireworks-ai openai

Set API key

import os
os.environ["FIREWORKS_API_KEY"] = "fw_..."  # from https://fireworks.ai/api-keys

Basic chat completion

from openai import OpenAI

client = OpenAI(
    base_url="https://api.fireworks.ai/inference/v1",
    api_key=os.environ["FIREWORKS_API_KEY"],
)

response = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain gradient descent in two sentences."},
    ],
    max_tokens=256,
    temperature=0.7,
)
print(response.choices[0].message.content)

Inference

Chat completions

Endpoint: POST https://api.fireworks.ai/inference/v1/chat/completions

from openai import OpenAI

client = OpenAI(
    base_url="https://api.fireworks.ai/inference/v1",
    api_key=os.environ["FIREWORKS_API_KEY"],
)

response = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    messages=[
        {"role": "user", "content": "Write a Python quicksort function."},
    ],
    max_tokens=512,
    temperature=0.0,
)
print(response.choices[0].message.content)

Streaming

stream = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    messages=[{"role": "user", "content": "Explain transformers."}],
    stream=True,
    max_tokens=512,
)
for chunk in stream:
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="", flush=True)

Function calling / tool use

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get current weather for a location",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {"type": "string", "description": "City name"},
                    "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
                },
                "required": ["location"],
            },
        },
    }
]

response = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    messages=[{"role": "user", "content": "What is the weather in San Francisco?"}],
    tools=tools,
    tool_choice="auto",
)
tool_call = response.choices[0].message.tool_calls[0]
print(tool_call.function.name, tool_call.function.arguments)

Structured output / JSON mode

response = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    messages=[{"role": "user", "content": "List 3 planets with mass and diameter."}],
    response_format={"type": "json_object"},
    max_tokens=512,
)
import json
data = json.loads(response.choices[0].message.content)

For strict schema enforcement, use response_format with a JSON schema:

response = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    messages=[{"role": "user", "content": "Extract name and age from: John is 30."}],
    response_format={
        "type": "json_object",
        "schema": {
            "type": "object",
            "properties": {
                "name": {"type": "string"},
                "age": {"type": "integer"},
            },
            "required": ["name", "age"],
        },
    },
)

Vision (multimodal)

response = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p2-11b-vision-instruct",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Describe this image."},
                {
                    "type": "image_url",
                    "image_url": {"url": "https://example.com/photo.jpg"},
                },
            ],
        }
    ],
    max_tokens=256,
)
print(response.choices[0].message.content)

Fine-Tuning

Fireworks supports three fine-tuning methods: Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Fine-Tuning (RL). All use LoRA adapters by default.

SFT -- Supervised Fine-Tuning

Data format (JSONL): Each line is a conversation with messages array:

{"messages": [{"role": "system", "content": "You are a coding assistant."}, {"role": "user", "content": "Write a Python hello world."}, {"role": "assistant", "content": "print('Hello, world!')"}]}
{"messages": [{"role": "user", "content": "What is 2+2?"}, {"role": "assistant", "content": "4"}]}

Create SFT job via API:

import requests

url = "https://api.fireworks.ai/v1/accounts/{account_id}/supervisedFineTuningJobs"
headers = {
    "Authorization": f"Bearer {os.environ['FIREWORKS_API_KEY']}",
    "Content-Type": "application/json",
}
payload = {
    "displayName": "my-sft-job",
    "model": "accounts/fireworks/models/llama-v3p3-70b-instruct",
    "dataset": "accounts/{account_id}/datasets/{dataset_id}",
    "hyperparameters": {
        "epochs": 3,
        "learning_rate": 1e-4,
        "batch_size": 8,
        "lora_rank": 16,
    },
}
response = requests.post(url, headers=headers, json=payload)
job = response.json()
print(f"Job ID: {job['name']}")

Monitor job:

job_url = f"https://api.fireworks.ai/v1/{job['name']}"
status = requests.get(job_url, headers=headers).json()
print(f"State: {status['state']}, Progress: {status.get('progress', 'N/A')}")

DPO -- Direct Preference Optimization

Data format (JSONL): Each line has chosen and rejected conversations:

{"chosen": [{"role": "user", "content": "Explain ML"}, {"role": "assistant", "content": "Machine learning is..."}], "rejected": [{"role": "user", "content": "Explain ML"}, {"role": "assistant", "content": "ML is complicated..."}]}

Create DPO job:

url = "https://api.fireworks.ai/v1/accounts/{account_id}/dpoFineTuningJobs"
payload = {
    "displayName": "my-dpo-job",
    "model": "accounts/fireworks/models/llama-v3p3-70b-instruct",
    "dataset": "accounts/{account_id}/datasets/{dataset_id}",
    "hyperparameters": {
        "epochs": 2,
        "learning_rate": 5e-5,
        "beta": 0.1,
    },
}
response = requests.post(url, headers=headers, json=payload)

RL -- Reinforcement Fine-Tuning

Reinforcement fine-tuning uses a reward model or reward function to optimize the base model. Jobs run on on-demand GPUs and are billed at GPU-hour rates.

Create RL job:

url = "https://api.fireworks.ai/v1/accounts/{account_id}/reinforcementFineTuningJobs"
payload = {
    "displayName": "my-rl-job",
    "baseModel": "accounts/fireworks/models/llama-v3p3-70b-instruct",
    "rewardModel": "accounts/{account_id}/models/{reward_model_id}",
    "dataset": "accounts/{account_id}/datasets/{dataset_id}",
}
response = requests.post(url, headers=headers, json=payload)

Fine-tuning pricing

Model Size	SFT ($/hr)	DPO ($/hr)
Up to 16B	$0.50	$1.00
16B - 80B	$3.00	$6.00
80B - 300B	$6.00	$12.00
300B+	$10.00	$20.00

RL fine-tuning is billed at on-demand GPU rates.

On-Demand GPU Deployments

Dedicated GPU deployments provide predictable latency, no rate limits, and support for custom/fine-tuned models. Billed per GPU-second.

Create deployment

import requests

url = "https://api.fireworks.ai/v1/accounts/{account_id}/deployments"
headers = {
    "Authorization": f"Bearer {os.environ['FIREWORKS_API_KEY']}",
    "Content-Type": "application/json",
}
payload = {
    "displayName": "my-llama-deployment",
    "model": "accounts/fireworks/models/llama-v3p3-70b-instruct",
    "deploymentShape": "fast",  # Options: fast, throughput, minimal
    "minReplicaCount": 1,
    "maxReplicaCount": 4,
}
response = requests.post(url, headers=headers, json=payload)
deployment = response.json()

Deployment shapes

Shape	Optimized For	Use Case
fast	Lowest latency	Real-time chat, interactive apps
throughput	Maximum tokens/sec	Batch processing, high volume
minimal	Lowest cost	Development, testing

GPU options and pricing

GPU	VRAM	Price/GPU/hr
A100 80GB	80 GB	$2.90
H100 80GB	80 GB	$4.00
H200 141GB	141 GB	$6.00
B200 180GB	180 GB	$9.00

Manage deployments

# List deployments
deployments = requests.get(
    f"https://api.fireworks.ai/v1/accounts/{{account_id}}/deployments",
    headers=headers,
).json()

# Scale deployment
requests.patch(
    f"https://api.fireworks.ai/v1/{deployment['name']}",
    headers=headers,
    json={"minReplicaCount": 2, "maxReplicaCount": 8},
)

# Delete deployment
requests.delete(
    f"https://api.fireworks.ai/v1/{deployment['name']}",
    headers=headers,
)

Query your deployment

Once deployed, query using the same OpenAI-compatible API but with your deployment's model ID:

response = client.chat.completions.create(
    model="accounts/{account_id}/deployments/{deployment_id}",
    messages=[{"role": "user", "content": "Hello!"}],
)

Embeddings

Endpoint: POST https://api.fireworks.ai/inference/v1/embeddings

Supported embedding models

Model ID	Dimensions
nomic-ai/nomic-embed-text-v1.5	768
nomic-ai/nomic-embed-text-v1	768
thenlper/gte-large	1024
WhereIsAI/UAE-Large-V1	1024

Generate embeddings

response = client.embeddings.create(
    model="nomic-ai/nomic-embed-text-v1.5",
    input=["Machine learning is a subset of AI.", "Deep learning uses neural networks."],
)
for i, emb in enumerate(response.data):
    print(f"Embedding {i}: {len(emb.embedding)} dimensions")

Model Selection

Popular serverless models

Model	Model ID	Params	Context	Tier
Llama 3.3 70B Instruct	accounts/fireworks/models/llama-v3p3-70b-instruct	70B	131K	>16B
Llama 3.2 11B Vision	accounts/fireworks/models/llama-v3p2-11b-vision-instruct	11B	128K	4-16B
Llama 3.2 3B Instruct	accounts/fireworks/models/llama-v3p2-3b-instruct	3B	128K	<4B
Qwen 2.5 72B Instruct	accounts/fireworks/models/qwen2p5-72b-instruct	72B	32K	>16B
Qwen3 Coder 480B A35B	accounts/fireworks/models/qwen3-coder-480b-a35b-instruct	480B (35B active)	262K	MoE
DeepSeek V3	accounts/fireworks/models/deepseek-v3-0324	671B (37B active)	164K	MoE
Mixtral 8x7B Instruct	accounts/fireworks/models/mixtral-8x7b-instruct	46B (12B active)	32K	MoE 0-56B
Mixtral 8x22B Instruct	accounts/fireworks/models/mixtral-8x22b-instruct	141B (39B active)	65K	MoE 56-176B

Serverless pricing by tier

Tier	Price per 1M tokens
< 4B params	$0.10
4B - 16B params	$0.20
> 16B params	$0.90
MoE 0 - 56B params	$0.50
MoE 56B - 176B params	$1.20

Model selection guide

Use Case	Recommended Model
General chat / instruction following	llama-v3p3-70b-instruct
Code generation	qwen3-coder-480b-a35b-instruct
Vision / multimodal	llama-v3p2-11b-vision-instruct
Cost-sensitive workloads	llama-v3p2-3b-instruct
Reasoning / complex tasks	deepseek-v3-0324
Fast MoE inference	mixtral-8x7b-instruct

CLI (firectl)

Installation

# macOS / Linux (Homebrew)
brew tap fw-ai/firectl && brew install firectl

# Install script
curl -sSL https://cli.fireworks.ai/install.sh | bash

# Windows (Chocolatey)
choco install firectl

# Verify
firectl version

# Upgrade
firectl upgrade

Authentication

firectl signin      # Interactive login
firectl whoami      # Show current account

Model management

# Upload a custom model
firectl model create my-model /path/to/model/weights

# List models
firectl model list

# Get model details
firectl model get accounts/{account_id}/models/my-model

# Delete model
firectl model delete accounts/{account_id}/models/my-model

Deployment management

# Create on-demand deployment
firectl deployment create accounts/fireworks/models/llama-v3p3-70b-instruct \
    --display-name "prod-llama"

# List deployments
firectl deployment list

# Scale deployment
firectl deployment scale {deployment_id} \
    --min-replica-count 2 --max-replica-count 8

# Delete deployment
firectl deployment delete {deployment_id}

Fine-tuning via CLI

# Create SFT job
firectl supervised-fine-tuning-job create my-sft-job \
    --model accounts/fireworks/models/llama-v3p3-70b-instruct \
    --dataset accounts/{account_id}/datasets/my-dataset

# Create RL fine-tuning job
firectl reinforcement-fine-tuning-job create my-rl-job \
    --base-model accounts/fireworks/models/llama-v3p3-70b-instruct \
    --reward-model accounts/{account_id}/models/my-reward-model

# Monitor jobs
firectl fine-tuning-job list
firectl fine-tuning-job get my-sft-job

# Stop / resume
firectl fine-tuning-job stop my-sft-job
firectl fine-tuning-job resume my-sft-job

OpenAI Compatibility

Fireworks AI is a drop-in replacement for the OpenAI Python SDK. Change base_url and api_key -- all existing code works unchanged.

Using the OpenAI SDK

from openai import OpenAI

client = OpenAI(
    base_url="https://api.fireworks.ai/inference/v1",
    api_key=os.environ["FIREWORKS_API_KEY"],
)

# Chat completions -- same API as OpenAI
response = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    messages=[{"role": "user", "content": "Hello!"}],
)

# Streaming -- same API
stream = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    messages=[{"role": "user", "content": "Hello!"}],
    stream=True,
)

# Embeddings -- same API
embeddings = client.embeddings.create(
    model="nomic-ai/nomic-embed-text-v1.5",
    input=["text to embed"],
)

Environment variable approach

export OPENAI_API_BASE="https://api.fireworks.ai/inference/v1"
export OPENAI_API_KEY="fw_..."

Then use the OpenAI SDK without any code changes.

Fireworks-specific parameters

Fireworks adds context_length_exceeded_behavior to control what happens when prompt + max_tokens exceeds the model's context window:

response = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    messages=[{"role": "user", "content": "..."}],
    max_tokens=512,
    extra_body={"context_length_exceeded_behavior": "truncate"},  # or "error"
)

Using the native Fireworks SDK

import fireworks.client

fireworks.client.api_key = os.environ["FIREWORKS_API_KEY"]

response = fireworks.client.ChatCompletion.create(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

Cost Optimization

Prompt caching

Fireworks automatically caches repeated prompt prefixes. No configuration needed -- identical prefixes across requests reuse cached KV states, reducing both latency and cost.

Best practices for prompt caching:

• Place static system prompts at the beginning of messages
• Keep dynamic content at the end
• Use consistent system prompts across requests

Batch inference

Batch API provides up to 50% cost savings for non-latency-sensitive workloads:

# Prepare batch file (JSONL)
# Each line: {"custom_id": "req-1", "method": "POST", "url": "/v1/chat/completions", "body": {...}}

# Upload batch file
batch_file = client.files.create(
    file=open("batch_requests.jsonl", "rb"),
    purpose="batch",
)

# Create batch job
batch = client.batches.create(
    input_file_id=batch_file.id,
    endpoint="/v1/chat/completions",
    completion_window="24h",
)
print(f"Batch ID: {batch.id}, Status: {batch.status}")

# Check status
batch_status = client.batches.retrieve(batch.id)
print(f"Status: {batch_status.status}")

Cost reduction strategies

Strategy	Savings	How
Use smaller models	50-90%	llama-v3p2-3b-instruct at $0.10/M tokens vs 70B at $0.90/M
Batch API	~50%	Async processing for non-real-time workloads
Prompt caching	20-40%	Consistent system prompts, static prefixes
MoE models	30-50%	Mixtral/DeepSeek: large capacity, smaller active params
On-demand deployments	Variable	Predictable pricing at scale, no per-token markup
Reduce max_tokens	10-30%	Set realistic output limits

Common Issues

Problem	Solution
401 Unauthorized	Check FIREWORKS_API_KEY is set and valid. Get key from https://fireworks.ai/api-keys
Model not found	Use full model ID: accounts/fireworks/models/{model_name}
Context length exceeded	Reduce input or set context_length_exceeded_behavior: "truncate"
Rate limited (serverless)	Switch to on-demand deployment for no rate limits
Slow cold start on deployment	Set minReplicaCount >= 1 to keep replicas warm
Fine-tuning job stuck	Check dataset format matches expected JSONL schema. Use firectl fine-tuning-job get
Tool calls not working	Use models that support function calling (Llama 3.3, Qwen, DeepSeek V3)
JSON mode returns invalid JSON	Use response_format with explicit schema for strict enforcement
Streaming usage stats missing	Upgrade openai SDK to >= 1.6.1. Usage is in the final stream chunk
Deployment not scaling	Check maxReplicaCount is set high enough. Review deployment shape

Resources

• Documentation: https://docs.fireworks.ai
• API Reference: https://docs.fireworks.ai/api-reference/introduction
• Dashboard: https://fireworks.ai/dashboard
• Model Catalog: https://fireworks.ai/models
• Pricing: https://fireworks.ai/pricing
• firectl CLI: https://docs.fireworks.ai/tools-sdks/firectl/firectl
• OpenAI Compatibility: https://docs.fireworks.ai/tools-sdks/openai-compatibility
• Fine-Tuning Guide: https://docs.fireworks.ai/fine-tuning/fine-tuning-models
• Cookbook (GitHub): https://github.com/fw-ai/cookbook
• Status Page: https://status.fireworks.ai