Provocative — Green Inference

Drop-in compatible

Change one line. Keep your stack.

If you already call the OpenAI API, you already know ours. Point your base URL at Provocative and keep everything else — same SDK, same request shape, same streaming.

⚡
Consistent, not just fast
Our p95 latency beats most providers' median. Predictable under load — the part that actually breaks in production.
🛡️
Doesn't fall over
Zero errors across thousands of concurrent requests while marketplace providers shed 70%+ of traffic.
🌱
green, same code
Direct-air capture at the rack. Going green costs you nothing in latency or lines of code.

inference.py

# Same OpenAI SDK — just a new base URL
from openai import OpenAI

client = OpenAI(
    base_url="https://inference.provocative.earth/v1",
    api_key="pk-prov-...",
)

resp = client.chat.completions.create(
    model="qwen3.6-35b",
    messages=[{"role": "user",
               "content": "Why run green?"}],
    stream=True,
)

terminal

# Or plain curl — your key is a bearer token
curl https://inference.provocative.earth/v1/chat/completions \
  -H "Authorization: Bearer pk-prov-..." \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3.6-35b",
       "messages": [{"role": "user", "content": "Why run green?"}]}'

Measured, not marketed

Aggregate throughput at peak load.

Qwen3.6-35B, 64 concurrent requests, measured from a neutral cloud vantage. Higher is better — and ours kept a 0% error rate the whole way.

Provocativeown hardware · 0% errors

2237

OR → AtlasCloudmarketplace · 77% errors

1736

OR → automarketplace · 0% errors

1716

OR → Parasailmarketplace · 74% errors

1327

tok/s aggregate @ c=64; throughput counts successful requests only. Parasail and AtlasCloud dropped 74–77% of requests at this concurrency — ours dropped none. No LPU or specialty-hardware provider serves Qwen3.6, so this is GPU versus GPU.

The beta model

One open model to start: Qwen3.6-35B.

We're launching the beta with a single, highly capable open model — a sparse mixture-of-experts served at native BF16 (no silent quantization) on owned A100 and Blackwell hardware, with Qwen's reasoning and tool-call parsers configured the way the model expects. One endpoint, tuned hard, measured in public. More coming through the beta.

qwen3.6-35b

MoE · 35B / 3B active

262k

context

196ms

median TTFT @ c=64

BF16

native precision

How to connect

Self-serve, OpenAI-compatible.

Point any OpenAI-compatible client — or plain curl — at our endpoint. Send your API key as a bearer token and set the model to qwen3.6-35b. Qwen runs with thinking off by default for fast, clean completions — pass chat_template_kwargs to control it.

Base URLhttps://inference.provocative.earth/v1

AuthAuthorization: Bearer <YOUR_API_KEY>

Model field"model": "qwen3.6-35b"

quickstart.sh

# Chat completion via curl — thinking disabled for speed
curl https://inference.provocative.earth/v1/chat/completions \
  -H "Authorization: Bearer $PROVOCATIVE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.6-35b",
    "messages": [{"role": "user", "content": "Why run green?"}],
    "chat_template_kwargs": {"enable_thinking": false}
  }'

Why green

Bold pink. Deep green. No contradiction.

Most AI runs on grid power and vents heat into the world. We built it backwards: our datacenters integrate direct-air carbon capture, so the same compute that answers your prompt also pulls CO₂ back out of the atmosphere. Same tokens, a smaller footprint — no offsets, no accounting tricks. It's in the hardware.

DAC

direct-air capture integrated at the rack

Owned

our metal, our power — not rented cloud

No offsets

green at the source, not on paper

Same API

green costs you nothing in code

Free beta

Be first on the green endpoint.

The first 100 users each get 1,000,000 free tokens on qwen3.6-35b. Drop your email and we'll send your API key — no payment step, no card.

No spam — just your key and beta updates. See our privacy policy.

How the beta works

Step 1

Join the waitlist

One field, one click. We onboard in small batches.

Step 2

Get your free API key

When your spot opens, your key lands in your inbox with 1,000,000 free tokens preloaded. No payment, no card.

Step 3

Point your SDK at us

Set base_url to inference.provocative.earth/v1, model qwen3.6-35b — same OpenAI SDK, zero other changes.

Change one line. Keep your stack.

Consistent, not just fast

Doesn't fall over

green, same code