Superion S1 — calibrated decisions in one call

#1 of 214 on S1MB English v1 — Task Average 75.06 · Borda 89.82 (12 Nov 2026). See the leaderboard →

Send a text and ask typed questions. Get back one calibrated probability per question — in one call, with no text generation, in well under a second. The answer goes straight into an if.

🔑 Get your free API key · ▶ Try the live demo · 📖 Docs · 🏭 Custom model for your industry


Ranking	#1 of 214 on S1MB English v1 · Task Average 75.06 · Borda 89.82 (S1MB leaderboard, 12 Nov 2026)
Question types	`noul` (yes/no) · `choice` (2–255 options) · `score` (2–16 ordered levels)
Output	calibrated probabilities you can threshold — no parsing, no hallucinated labels
Speed	Sub-second per request, many questions per request
Context	up to 32K tokens per question
Price	Free: 20 M tokens/day per key, forever · Prepaid: 0.042 USD per million input tokens, output free · Contracts: dedicated instances
Model	built on Gemma 4 26B-A4B, INT4 experts · served as a hosted API

This repository is the model card. Superion S1 is available as an API only — no weights are distributed.

Why teams use it

Most "LLM as a classifier" pipelines ask a chat model to write an answer, then parse it, then hope the confidence means something. Superion S1 skips all of that:

Probabilities, not prose. Every answer is a number between 0 and 1. Set your threshold once, route the rest to a human.
Many decisions, one call. Intent, urgency, refund, sentiment — ask them together about the same text and pay for the text once.
Deterministic within a pinned model version. The same request to the same pinned version returns the same probabilities — reproducible automations and audits.
Standard System One format. choice, score, confidence, legend and a request_id on every response; existing System One clients only change the base URL and key.
Built for volume. Sub-second, no output tokens to pay for, and a Free plan big enough for real workloads.
Proven in public. #1 out of 214 models on the open S1MB benchmark (public leaderboard, 12 Nov 2026).

Quickstart

Get a free key at punkytigerlabs.com/s1 (email only, no card).

Python

import os, requests

r = requests.post(
    "https://api.punkytigerlabs.com/v1/systemone",
    headers={"Authorization": "Bearer " + os.environ["S1_KEY"]},
    timeout=120,
    json={
        "model": "superion-s1-v6-int4e-vllm-kv1@sha256:e6dc49dc",
        "state": "Message: The laptop arrived broken. I want my money back now, this is unacceptable.",
        "questions": {
            "refund":    {"type": "noul",   "instructions": "Does the customer ask for a refund or a replacement?",
                          "criteria": {"true": "yes", "false": "no"}},
            "sentiment": {"type": "choice", "instructions": "Overall sentiment of the message.",
                          "criteria": {"pos": "positive", "neg": "negative", "neu": "neutral"}},
            "urgency":   {"type": "score",  "instructions": "How urgent is a reply?",
                          "criteria": ["not urgent", "somewhat urgent", "very urgent"]},
        },
    },
)
r.raise_for_status()
answers = r.json()["answers"]
if answers["refund"]["noul"] > 0.9:
    print("→ refunds queue")

curl

curl https://api.punkytigerlabs.com/v1/systemone \
  -H "Authorization: Bearer $S1_KEY" -H "Content-Type: application/json" \
  -d '{"model": "superion-s1-v6-int4e-vllm-kv1@sha256:e6dc49dc",
       "state": "Message: The laptop arrived broken. I want my money back now, this is unacceptable.",
       "questions": {
         "refund":    {"type": "noul",   "instructions": "Does the customer ask for a refund or a replacement?", "criteria": {"true": "yes", "false": "no"}},
         "sentiment": {"type": "choice", "instructions": "Overall sentiment of the message.", "criteria": {"pos": "positive", "neg": "negative", "neu": "neutral"}},
         "urgency":   {"type": "score",  "instructions": "How urgent is a reply?", "criteria": ["not urgent", "somewhat urgent", "very urgent"]}}}'

Response (real, 12 Nov 2026, v6, rounded)

{
  "model": "superion-s1-v6-int4e-vllm-kv1@sha256:e6dc49dc",
  "answers": {
    "refund":    {"type": "noul",   "noul": 0.9998},
    "sentiment": {"type": "choice", "choice": "neg", "confidence": 0.9661, "probabilities": {"pos": 0.0003, "neg": 0.9941, "neu": 0.0056}},
    "urgency":   {"type": "score",  "score": 1.9393, "confidence": 0.7917, "probabilities": {"0": 0.0000, "1": 0.0607, "2": 0.9393},
                  "legend": {"0": "not urgent", "1": "somewhat urgent", "2": "very urgent"}}
  },
  "usage": {"input_tokens": 135, "output_tokens": 0},
  "meta": {"request_id": "63a6718e-ab64-404b-92f3-fdba826d0dfe", "mode": "live", "latency_ms": 103}
}

Three decisions, 135 billed tokens. At the Prepaid price that is ≈ 5.7 USD per million tickets triaged — and the Free plan covers roughly 148,000 such requests per day.

Type	`criteria`	Answer
`noul`	`{"true": "...", "false": "..."}`	`noul`: probability of true
`choice`	`{option_id: description, ...}` (2–255)	`choice`, `confidence` and `probabilities` per option id
`score`	`["lowest", ..., "highest"]` (2–16 levels)	`score` (expected level), `confidence`, `probabilities` per level index and `legend`

Full reference, limits and error codes: punkytigerlabs.com/s1/docs.

Use cases

You run…	Ask Superion S1…
AI agents	Is this action safe to execute? Did the user actually confirm? Is the tool output an error? — a calibrated gate before every irreversible step
Support & ops	Intent? Urgency? Refund? Escalate? — full ticket triage in one request
User-generated content	Harassment? Spam? Self-harm risk? PII? — moderation with thresholds you can trust
Sales & marketing	Is this lead a fit? Buying intent 1–5? Reply needed? — score every inbound email and form
LLM platforms	Which model should answer this? Is this answer grounded? — a fast, cheap router and judge
Search & RAG	Is this passage relevant to the query? — relevance scoring as a reranking signal

🏭 Custom & industry models

Superion S1 is our general model. We also build custom models and our own specialized models per industry or objective — tuned to your taxonomy, your decision criteria and your data, and served on a dedicated instance with the same API:

Fintech & risk — transaction review, KYC/AML alert triage, dispute classification
Healthcare — intake routing and triage support for operational workflows
E-commerce & customer care — returns, refunds, intent and escalation at scale
Legal & compliance — clause and policy checks, contract review flags, regulatory screening
Trust & safety — moderation tuned to your community guidelines
Agent routing & guardrails — the decision layer for your agent stack

Same one-call interface, same calibrated probabilities, your own private model and capacity. Need a model for your industry? Talk to us → alvaro.lopez@punkytigerlabs.com

Pricing

Plan	Price	Limits
Free	0 USD — permanent, no card	20 M input tokens per day per key
Prepaid	0.042 USD / M input tokens · output always free	240 req/min · 80,000 tokens/min · credits never expire
Contract / Custom	let's talk	dedicated instance, priority capacity, custom or industry model

Only successful (200) responses are billed. Sign up →

Models served

ID	Status
`superion-s1-v6-int4e-vllm-kv1@sha256:e6dc49dc`	Recommended (the #1 model)
`superion-s1-v5-int4e-vllm-kv1@sha256:47f7571c`	Available

IDs are immutable; retirement comes with at least 90 days' notice. GET https://api.punkytigerlabs.com/v1/models lists what is live.

Benchmark

S1MB is an open benchmark for System One decision models: 137 English benchmarks across yes/no, choice and score decisions. Superion S1 v6 ranks #1 of 214 on both of its headline metrics — Task Average 75.06 and Borda 89.82 — on the public leaderboard as of 12 Nov 2026. Rankings move as new models are submitted; the live table is at hotchpotch/s1mb-result (row punkytiger__superion_s1).

Model tree for PunkyTiger/Superion-S1

Base model

google/gemma-4-26B-A4B

Finetuned

google/gemma-4-26B-A4B-it