Coming from Fireworks AIKeep Fireworks AI. Add every other lab.
Same OpenAI-compatible mechanics as Fireworks' serverless endpoint — shorter model ids, one prepaid per-token balance, and frontier closed models next to your open ones.
api.fireworks.ai/inference/v1
www.ninjachat.ai/api/v1accounts/fireworks/models/… becomes the slug.
accounts/fireworks/models/deepseek-v3p2→deepseek-v3.2
accounts/fireworks/models/qwq-32b→qwq-32b
accounts/fireworks/models/kimi-k2-instruct→kimi-k2.6from openai import OpenAI
client = OpenAI(
base_url="https://api.fireworks.ai/inference/v1",
api_key="fw_...",
)
resp = client.chat.completions.create(
model="accounts/fireworks/models/deepseek-v3p2",
messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.choices[0].message.content)from openai import OpenAI
client = OpenAI(
base_url="https://www.ninjachat.ai/api/v1", # ← changed
api_key="nj_sk_YOUR_KEY", # ← changed
)
resp = client.chat.completions.create(
model="deepseek-v3.2",
messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.choices[0].message.content)What joining NinjaChat actually adds.
FLUX 2
Imagen 4
Nano Banana
Recraft V3
Seedance 2
GPT-5.4
Claude Opus
Kling 2.6Then these.
Create a key and add credits
Get an nj_sk_ key at /developers and add a credit pack. Balances are prepaid, and failed generations are refunded.
verify your key works# Sanity check: list every model your key can call (no auth needed) curl https://www.ninjachat.ai/api/v1/models # First authenticated request curl https://www.ninjachat.ai/api/v1/chat/completions \ -H "Authorization: Bearer nj_sk_YOUR_KEY" \ -H "Content-Type: application/json" \ -d '{"model": "gpt-5", "messages": [{"role": "user", "content": "Say hi"}]}'Drop the accounts/fireworks/models/ prefix
Fireworks ids carry the deployment path. NinjaChat slugs are just the model: accounts/fireworks/models/deepseek-v3p2 → deepseek-v3.2. Mapping table below; anything not listed can be checked live via GET /v1/models.
Verify streaming and tool calling
Same SDK code. Streaming arrives as SSE deltas, and tool_calls round-trip with role: "tool" messages.
streaming + tools, unchanged sdk# Streaming stream = client.chat.completions.create( model="claude-sonnet-4.6", messages=[{"role": "user", "content": "Write a haiku about ninjas"}], stream=True, ) for chunk in stream: if chunk.choices and chunk.choices[0].delta.content: print(chunk.choices[0].delta.content, end="", flush=True) # Tool calling resp = client.chat.completions.create( model="gpt-5", messages=[{"role": "user", "content": "Weather in Tokyo?"}], tools=[{ "type": "function", "function": { "name": "get_weather", "parameters": { "type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"], }, }, }], ) call = resp.choices[0].message.tool_calls[0] # ...run your function, then send the result back: followup = client.chat.completions.create( model="gpt-5", messages=[ {"role": "user", "content": "Weather in Tokyo?"}, resp.choices[0].message, {"role": "tool", "tool_call_id": call.id, "content": "22°C, clear"}, ], )
Model mapping — Fireworks ids to NinjaChat slugs
What doesn't come along.
No /v1/embeddings and no moderations endpoint
NinjaChat's v1 API covers chat, images, video, and search. If your pipeline embeds documents or calls a moderation endpoint, keep those calls on your current provider — only the completion traffic needs to move.
60 requests/min default rate limit
Every API key gets 60 requests per minute (video submissions are throttled harder because each one is a long-running job). 429 responses include Retry-After. Need more? Contact us from the console.
Prepaid credits, not postpaid billing
You buy a credit pack up front instead of getting a surprise invoice. Failed generations are automatically refunded, and you can set monthly spend limits per account, key, or project.
Metered $/MTok chat pricing
Chat bills per token — published $/MTok input and output rates for every model on GET /v1/models (cache-read and long-context tiers included). Requests preauthorize an estimated maximum and settle to actual usage; a typical request runs from ≈$0.002 on open models to ≈$0.05 on Claude Opus (gpt-5.4-pro ≈$0.33). No length guard — long context just needs balance.
No dedicated deployments, fine-tuning, or FireAttention tuning
Fireworks' on-demand deployments, LoRA fine-tuning, and performance-tuned dedicated instances have no equivalent here — NinjaChat is serverless multi-provider inference. Keep custom-model workloads on Fireworks and migrate the standard-model traffic.
No embeddings endpoint
Fireworks serves embeddings models; NinjaChat doesn't. Point an embeddings-only client at Fireworks and your completions client at NinjaChat — the OpenAI SDK supports multiple client instances with different base_urls.
FAQ
Is the migration really just two lines?
For the client, yes: base_url and api_key. Model ids also shrink from accounts/fireworks/models/deepseek-v3p2 to deepseek-v3.2 — a find-and-replace, with the table on this page as the reference.
Does grammar/JSON mode survive the move?
OpenAI-style response_format json_object and json_schema are enforced server-side. Fireworks' custom grammar mode (GBNF) has no equivalent — express constraints as a JSON schema instead.
What's the latency story compared to Fireworks?
NinjaChat routes to upstream providers rather than running its own FireAttention stack, so raw tokens/sec on open models can be lower than a tuned Fireworks deployment. What you gain is one key across open + closed models, image, and video. Benchmark your workload with the free playground before committing.
Can I keep fine-tuned models?
No — NinjaChat has no fine-tuning or custom-checkpoint hosting. Standard model traffic migrates; custom checkpoints stay on Fireworks.