Qwen3.6 35B Decide, decisions instead of generated text
Qwen3.6-35B-A3B from Alibaba's Qwen team, served to answer typed questions: send a state and your questions, get back labels, scores and probabilities. Each answer is one letter-constrained token, so only input tokens bill.
- Choice, score, yes or no
- Multilingual
- Input-only billing
- Apache 2.0
Three tiers, per token.
Per 1M tokens
| Tier | Input |
|---|---|
| Sync | $0.05 |
| AsyncEarly access | $0.04 |
Async is 20% off Sync.
Decisions meter input tokens only. Each answer is a single constrained token, so there is no output rate.
openrelay/qwen3.6-35b-decideFull catalog on the inference pricing page. Add a card and make a deposit to start.
import os
import requests
API_KEY = os.environ["OPENRELAY_API_KEY"] # an or_ key
resp = requests.post(
"https://inference.openrelay.inc/v1/decisions",
headers={"Authorization": f"Bearer {API_KEY}"},
json={
"model": "openrelay/qwen3.6-35b-decide",
"state": "We were billed twice for March. Refund it or we cancel.",
"questions": {
"department": {"type": "choice", "instructions": "Which team handles this?",
"criteria": ["billing", "technical", "other"]},
"churn_risk": {"type": "noul",
"instructions": "Does the user threaten to leave?"},
},
},
)
print(resp.json()["answers"])The spec sheet.
What the model is, what it takes in, and the surface it serves on.
- Model id
- openrelay/qwen3.6-35b-decide
- Base model
- Qwen3.6-35B-A3B: mixture of experts, 35B parameters, 3B active per token
- Question types
- choice (a label from your options), score (a level on your scale), noul (the probability of yes)
- How it answers
- One prompt per question with the options lettered A to Z; the answer is read from the letter probabilities
- Limits
- 64 questions per request, 26 options per question, 32 KB of state, read whole
- Endpoint
- POST /v1/decisions, online only (the Batch API does not accept it)
- License
- Apache 2.0, open weights
Where Qwen3.6 35B Decide earns its place.
What this model is actually for, versus the rest of the catalog.
Typed answers, nothing to parse
Each answer is a field: a label with its probabilities, a score, or the probability of yes. There is no generated text to parse or validate.
Every question reads the whole state
Nothing is truncated: each question sees the full state, up to 32 KB. Only input tokens bill, and ten questions about one ticket bill the ticket ten times, in one round trip.
Measure before you route
The probabilities are zero-shot and uncalibrated, and instruct models of this kind tend to be over-confident. Validate on a labeled sample and set thresholds from it.
Qwen3.6 35B Decide, answered.
What is Qwen3.6 35B Decide?
Qwen3.6-35B-A3B, an Apache 2.0 mixture-of-experts model from Alibaba's Qwen team, served behind POST /v1/decisions. Each question is one prompt with lettered options, and the answer comes from the probabilities of those letters.
How is it billed?
Per input token, with no output rate. The state is part of every question's prompt, so usage.prompt_tokens is the sum over questions of the state, the instructions, the options and a short system prompt.
Which languages does it handle?
Qwen lists 201 languages for the model. The limits and the price are the same on every input.
Can I run decisions through the Batch API?
No. /v1/decisions is online only, and the Batch API does not accept it.
First request in five lines.
POST to inference.openrelay.inc/v1/decisions and run Qwen3.6 35B Decide per input token. No contract, no minimums.