Fireworks AI

Listed at https://fireworks.ai

Overall Rank #11 ⭐ Consider
❌ Proxy required | 🌍 International

💰 Token Pricing

TypePriceNote
Input Firefunction-v2: $0.90/M, Llama 3.3 70B: $0.90/M, DeepSeek-R1: $2.00/M per million tokens
Output Firefunction-v2: $0.90/M, Llama 3.3 70B: $0.90/M, DeepSeek-R1: $2.00/M per million tokens
💡 Free Credits:

🤖 Supported Models (100)

Llama 3.3 70BFirefunction-v2QWEN 2.5 72BDeepSeek-R1Mixtral 8x22BCodeLlama 70BLlama 3.1 405B

✨ Pros

  • Firefunction-v2 specifically optimized for Function Calling
  • Llama 3.3 70B at only $0.90/M tokens, very affordable
  • Moonshot engine for fast inference
  • Supports LoRA fine-tuning deployment
  • 100+ open-source models hosted

⚠️ Cons

  • ×China requires proxy
  • ×Models other than Firefunction-v2 less competitive
  • ×No free credits
  • ×Smaller company, long-term stability concern
  • ×Non-OpenAI-compatible models need extra adaptation

🎯 Best For

Function Calling-intensive apps; fast local inference; open-source model deployment testing

💰 Pricing & Plans

ModelInput ($/M)Cached ($/M)Output ($/M)Best For
Kimi K3$3.00$0.30$15.00Frontier reasoning & agents
Kimi K2.7 Code$0.95$0.19$4.00Coding and code agents
DeepSeek V4 Pro (0813)$1.74$0.145$3.48Deep reasoning, 1M context
DeepSeek V4 Flash (0731)$0.14$0.028$0.28Cheap high-volume tasks
GLM 5.2$1.40$0.14$4.40General chat + agentic use
Qwen 3.7 Plus$0.40$0.08$1.60Budget general-purpose
MiniMax M3$0.30$0.06$1.20Lightweight / budget apps
OpenAI GPT OSS 120B$0.15$0.015$0.60Open-weights generalist
Muse Glimmer 30B$0.35$0.04$1.50Multimodal / vision
Other open models$0.10–$1.20$0.10–$1.20By size: MoE ≤56B $0.50, dense >16B $0.90

🔧 API & Developer Experience

  • API Surface: OpenAI-compatible endpoint at https://api.fireworks.ai/inference/v1 — most OpenAI SDKs work by swapping base_url and api_key. Supports chat completions, streaming, function calling, structured outputs, vision and embeddings.
  • Signup & Credits: Fully self-serve at app.fireworks.ai; new accounts get $1 in free credits and postpaid billing with high serverless rate limits. No sales call required to start.
  • Serving Paths: Three serverless tiers: Standard (default), Priority (higher rate limits and throughput, ~1.5x price), and Fast (lower latency). US-only endpoints cost a 10% premium for data-residency needs.
  • Prompt Caching & Batch: Cached input tokens are billed about 90% cheaper (e.g. DeepSeek V4 Pro $0.145 vs $1.74 per M), and batch inference runs at 50% of serverless prices on input and output.
  • Fine-Tuning & Training: LoRA SFT/DPO, full-parameter SFT/DPO and RL fine-tuning; fine-tuned models are served at the base model's serverless price. Serverless Training API bills per token (no provisioning), Dedicated Training API per GPU hour.
  • On-Demand GPUs: H100 ($7/hr), H200 ($7), B200 ($10), B300 ($12), GB300 ($18) — per-second billing with no startup charge; region-restricted deployments at 1.5x.
  • Function Calling & Agents: Native tool-use with JSON schemas, plus Fireworks' Firefunction line of function-calling-optimized open models for agentic workloads.

🚀 High-Performance Open-Weight Serving & Moonshot Engine

Fireworks AI's core strength is high-performance serving of open-weight models: it hosts frontier open releases within days — DeepSeek V4 Pro 0813, Kimi K3, GLM 5.2, Qwen 3.7 Plus, MiniMax M3 and OpenAI's GPT OSS family all run on its Moonshot inference engine, which Fireworks claims delivers roughly 250% higher throughput and 50% lower latency than open-source inference stacks. The same platform sells on-demand GPUs (H100/H200/B200/B300/GB300) billed per second with no startup charge, plus a full fine-tuning stack (LoRA and full-parameter SFT/DPO with RL) where trained models are served at the base model's serverless price. That combination — latest open weights, fast serving, and GPU/fine-tuning on one bill — is the differentiator against pure API resellers.

🌐 China Access & Latency

Fireworks AI is a US company with no mainland China points of presence, so api.fireworks.ai is not optimized for China traffic. Direct connections from mainland typically require a proxy, VPN or overseas relay, and real-world latency is often 200–500ms+ depending on route quality — far worse than the sub-100ms of domestic providers. There is no China-specific pricing or billing. Workarounds: route through an OpenAI-compatible aggregator to reach the same models, or self-host the open-weight models (DeepSeek, Qwen, GLM, Kimi) on domestic GPU clouds — most of the catalog is open weights, so Fireworks is not the only host.