Fast, Faithful, Frontier System One Intelligence.
GS1 is a family of fast, uncertainty-aware, open-source decision models developed by BIGAI (Beijing Institute for General Artificial Intelligence). It computes native probability distributions over variable candidate sets for agent routing, verification, extraction, and control.
Code · Model · Documentation · Playground
- Fast - Native Runtime Decision Head.
- Produce Choice, Noul and Score distributions directly, with zero generated output tokens.
- A native multi-head cross-attention decision head queries state evidence using question and candidate representations.
- With CUDA Graph, GS1-0.8B reaches 12.51 ms P50 for one short question and 6.05× speedup over default inference for eight short questions.
- Faithful — multi-step RL with calibrated rewards
- Reinforcement learning optimizes sampled action probabilities with a Brier reward. RLCD
- Native probabilities and confidence support evidence gathering, deferral and explicit abstention choices.
- Frontier — Hybrid Architecture, Competitive Quality
- Causal Gated DeltaNet and Segmented Bidirectional Full Attention give candidates their complete comparison context while preserving immutable prefixes for reuse across questions.
- Carry-preserving Attention Residual lets sparse full-attention receivers query earlier representations while retaining the explicit skip-connection anchor. Architecture
- GS1-4B reaches 68.97 on the latest Decision Index 0.4.1 Public index. All GS1 sizes lie on the displayed parameter–performance frontier.
| Model | Decision Index (Public index) | Weights |
|---|---|---|
| GS1-0.8B | 56.12 | Hugging Face |
| GS1-2B | 62.22 | Hugging Face |
| GS1-4B | 68.97 | Hugging Face |
| GS1-9B | 70.45 | Hugging Face |
Exact stored counts and source identities are in the results.
These are the latest complete public-suite results (edition 0.3). All four models have 100% coverage across 43 text benchmarks, including 37 benchmarks that contribute to the Public index. Evaluation details.
Python 3.11+. Install the model folder's requirements.txt, or the corresponding PyTorch, Transformers 5.3.0, Accelerate, safetensors, Pydantic, TypeSafe SDK and Pillow dependencies. The gs1 repository is not required. Choose a release folder or its published Hub ID.
python -m pip install torch==2.9.1 torchvision==0.24.1 transformers==5.3.0 accelerate safetensors 'pydantic>=2.8,<3' typesafe-sdk==0.7.0 pillowimport torch
from transformers import AutoModel
model = (
AutoModel.from_pretrained(
"bigai/GS1-0.8B",
trust_remote_code=True,
)
.eval()
.to("cuda")
)
request = {
"state": "A customer was charged twice and requests an urgent refund.",
"questions": {
"team": {
"type": "choice",
"instructions": "Which team should handle this ticket?",
"criteria": {"billing": "Payments and refunds", "technical": "Software problems"},
},
"refund": {"type": "noul", "instructions": "Does the customer request a refund?"},
"urgency": {
"type": "score",
"instructions": "Rate the urgency of handling this ticket.",
"criteria": ["Routine", "Soon", "Immediate"],
},
},
}
with torch.autocast("cuda", dtype=torch.bfloat16, cache_enabled=False):
result = model.decide(**request, precision="bf16")
print(result["answers"])
print(result["usage"])result contains model, answers and usage. Every supplied candidate receives a probability. Full three-primitive example · Request JSON
The released chat_template.jinja renders the native GS1 state, question and options through the same serializer used by the compiler. model.decide prepares typed candidates and branch metadata. Native template usage
Linux/CUDA, Python 3.11+, PyTorch 2.9.1 and Transformers 5.3.0:
git clone https://github.com/bigai-nlco/GS1.git
cd GS1
uv sync --locked --extra serve --extra cudauv run gs1 serve --model bigai/GS1-0.8B --model-name GS1The service exposes POST /v1/systemone and GET /v1/models. Enable qualified CUDA Graph inference or use from gs1 import Engine for in-process inference. The runtime and Hub model share the model equations and typed output definitions.
Point the official TypeSafe Python SDK at the GS1 service started above. The client needs only the SDK:
python -m pip install typesafe-sdk==0.7.0from typesafe_sdk import Choice, Noul, Score, TypeSafeClient
with TypeSafeClient(
base_url="http://127.0.0.1:8000",
api_key="local",
model="GS1",
) as client:
result = client.system_one(
state="A customer was charged twice and requests an urgent refund.",
questions={
"team": Choice(
instructions="Which team should handle this ticket?",
criteria={"billing": "Payments and refunds", "technical": "Software problems"},
),
"refund": Noul(instructions="Does the customer request a refund?"),
"urgency": Score(
instructions="Rate the urgency of handling this ticket.",
criteria=["Routine", "Soon", "Immediate"],
),
},
)
print(result.choices["team"].choice)
print(result.choices["team"].probabilities)
print(result.nouls["refund"].noul)
print(result.scores["urgency"].score)api_key="local" satisfies the SDK's required argument; local GS1 requires no TypeSafe cloud credential. Use the API root as base_url and match model to --model-name. Runnable example · API and confidence
Latest inference measurements:
| Model | GS1 Default P50 | GS1 eager P50 | CUDA Graph P50 / P95 | Speedup | Graph allocated |
|---|---|---|---|---|---|
| GS1-0.8B | 54.10 ms | 54.60 ms | 12.51 / 13.22 ms | 4.32× | 4.07 GiB |
| GS1-2B | 53.95 ms | 54.34 ms | 18.64 / 18.69 ms | 2.89× | 10.60 GiB |
| GS1-4B | 72.87 ms | 73.84 ms | 35.17 / 35.24 ms | 2.07× | 23.68 GiB |
RTX 4090 reporting 48 GiB, BF16 computation, 256 tokens / one question, five warmups and 30 timed calls per group. These are complete in-process decision calls, including validation, prompt compilation, model execution, probabilities and typed responses; model loading, graph capture, HTTP and SDK transport are excluded. GS1 Default uses the standalone model's model.decide path; speedup compares its P50 with CUDA Graph at identical weights and inputs. Stored weight dtypes are preserved.
# From the source checkout; runs both 256/1,024-token and 1/8-question fixtures.
uv sync --locked --extra serve --extra cuda --extra eval
CUDA_VISIBLE_DEVICES=0 uv run python -m benchmark.latency \
--model bigai/GS1-0.8B --out runs/latency/GS1-0.8B.jsonAll shapes, P50/P95/P99, memory and methodology · Raw samples · Earlier service/IPC and replica measurements
The Query workspace shows sample Choice/Noul/Score answers, raw JSON and per-request round-trip latency. Example request.
cd playground
npm ci
npm run build
npm startOpen localhost:3000 with GS1 serving on port 8000. Use Query to edit state and questions, compare probabilities and inspect raw JSON. GS1 is the default backend. For hosted Jev, export TYPESAFE_API_KEY on the Node server to call the official TypeSafe API. Setup and controls
GS1 combines Qwen3.5 hybrid sequence mixing, segmented bidirectional full attention, carry-preserving Attention Residuals, and a native cross-attention decision head. Candidate meanings are supplied at inference time. Choice selects among them, Noul returns p(true), and Score returns the expected zero-based level. The head has no fixed candidate-count classifier; API and context budgets define the supported request sizes. Architecture and numerical contract
Model · RLCD · Inference · API · Evaluation
Code and released model weights use Apache-2.0.
If you find our work helpful, feel free to give us a cite.
@techreport{zheng2026gs1,
title={GS1: Fast, Faithful, Frontier System One Intelligence},
author={Zheng, Zilong and Liu, Yang and Qin, Shuhan and Lin, Ziyong and Li, Jiaqi and Jia, Zixia and Zhu, Song-Chun},
institution={BIGAI},
year={2026}
}








