animica-inference

Free keyless inference at animica.dev/v1

https://animica.dev/v1 is an OpenAI-compatible API (/v1/models, /v1/chat/completions). No API key required. Rate limit: 30 requests/min per IP.

Capacity is community-GPU-provided: each model carries a boolean serving flag in /v1/models. Check it first — a model with serving: false has no live worker at that moment.

1. Check what is serving

curl -s https://animica.dev/v1/models

Model ids (live list): kimi-k3 (default flagship), animica-chat, animica-chat-small, animica-chat-flagship, animica-knowledge (ENA, network-trained). Each entry includes "serving": true|false.

2. Chat completion — curl

curl -s -X POST https://animica.dev/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"kimi-k3","messages":[{"role":"user","content":"Say OK"}],"max_tokens":50}'

3. Chat completion — Python (openai SDK)

from openai import OpenAI

client = OpenAI(base_url="https://animica.dev/v1", api_key="none")  # keyless; SDK requires a placeholder
models = client.models.list()  # check the serving flag per model first
resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Say OK"}],
    max_tokens=50,
    stream=True,     # prefer streaming — completions take minutes, not seconds
    timeout=600,
)
text = "".join(chunk.choices[0].delta.content or "" for chunk in resp)

Honest error handling (important)

• Stub responses: when no external worker claims the job, the API can return HTTP 200 with a placeholder whose content begins with [distributed-aicf stub (verified live). Treat any response whose content starts with that marker as "model unavailable", not as an answer. Always check serving in /v1/models first, and validate the content.
• Latency: real completions typically take 1–3 minutes end-to-end. Use "stream": true and set client timeouts of 300 s or more (the gateway holds requests up to 600 s with SSE keepalives); do not retry aggressively — you will hit the 30 req/min/IP limit.
• Requests to non-serving models may 503, queue, or return the stub above — handle all three.
• Media endpoints (/v1/images, /v1/videos, /v1/audio) dispatch to GPU miners: submissions return 202 {status:"queued", job_id, poll_url} and queued jobs persist until a miner is online. A 503 no_media_miner means the dispatcher was unreachable and the submission did NOT enqueue — resubmit.
• /v1/embeddings is not available (404) — do not use this endpoint for embeddings.