GS1-0.8B (General System One)

Fast, Faithful, Frontier System One Intelligence.

Code · Quick Start · Evaluation · Playground

GS1 returns probabilities over user-defined candidates in a single decision pass. Use the YeS1 framework to run this model; standard text-generation pipelines do not implement its decision head.

GS1 is a family of fast, uncertainty-aware 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.

📌 Model Details

• Developed by: BIGAI (Beijing Institute for General Artificial Intelligence)
• Correspondence: Zilong Zheng (zlzheng@bigai.ai)
• Model Family: GS1-0.8B, GS1-2B, GS1-4B, GS1-9B
• Base Architecture: Qwen3.5 with hybrid Gated DeltaNet (GDN) and full-attention layers
• License: Apache 2.0
• Primary Task: Agentic decision making

🔥 GS1 Highlights

• Native Cross-Attention Decision Head: Computes probabilities directly over candidates. At fixed context, permuting head inputs preserves probabilities when candidate order is restored.
• Segmented Bidirectional Visibility: Combines causal GDN with segmented bidirectional full attention, allowing candidates to observe their comparison context while keeping shared prefixes reusable.
• Carry-Preserving Attention Residuals (AttnRes): Routes information from earlier representations while retaining the explicit skip-connection anchor.
• Typed Outputs: One probability distribution supports three decision modes:
  • Choice: Select a candidate and return every candidate's probability.
  • Noul: Return the probability that a statement is true.
  • Score: Return the expected zero-based ordinal level, $\sum_k k\pi_k$.

💡 Advanced Post-Training & Reliability

• Calibrated Reward Policy Optimization (CRPO): Uses sampled decision probabilities in a Brier-shaped penalty to align calibration rewards with the deployed policy.
• Learned Abstention: Optimizes accuracy and non-hallucination rates under Lagrangian constraints, making abstention an explicit action.
• Multi-Step RL: Extends CRPO to interaction trajectories with shared terminal feedback, including StarCraft II against the Very Hard computer opponent.
• Recursive Self-Improvement: Uses separate loops for data synthesis and infrastructure tuning.

🏆 Performance Benchmarks

Results on Decision Index 0.3 (Public):

Model	GS1 Score	Top Displayed Baseline	Gain
GS1-0.8B	56.15	53.79 (StartLux v1.1 0.8B)	+2.36
GS1-2B	62.22	60.64 (GLIDE 2B)	+1.58
GS1-4B	68.95	67.81 (StartLux v1.1 4B)	+1.14
GS1-9B	70.47	69.98 (StartLux v1.1 9B)	+0.49

Decision Index reports a composite score, not accuracy.

Evaluate your checkpoint

Run a separate AG News evaluation and save results with provenance:

uv run python -m evaluation --suites agnews \
  --model artifacts/GS1-0.8B --device cuda --precision bf16 \
  --out runs/evaluation/gs1-0.8b-agnews

This writes evaluation.json and accuracy_manifest.json. See the evaluation guide for other suites and Decision Index 0.3 preparation.

⚡ Inference Efficiency

GS1 replaces autoregressive answer generation with state encoding and decision-head evaluation. Shared execution can reuse state and head projections across questions.

On one RTX 4090, the trained GS1-2B artifact with BF16 compute and warmed reference CUDA graphs achieved 16.00 ms for one question (3.38×) and 357.76 ms for 32 questions (4.72×) against the single-branch reference baseline.

These are median engine timings for a fixed 2B workload, excluding compilation, HTTP overhead, and graph capture. They are not GS1-0.8B latency measurements. See measurement details.

🚀 Quick Start

Install and download

Use Python 3.13 and an NVIDIA CUDA environment for the GPU examples. The lockfile pins PyTorch 2.9.1 and Transformers 5.3.0.

git clone https://github.com/zilongzheng/YeS1.git
cd YeS1
uv sync --locked --extra train --extra sources

uv run python - <<'PY'
from huggingface_hub import model_info, snapshot_download

repo_id = "bigai/GS1-0.8B"
revision = model_info(repo_id).sha
snapshot_download(repo_id, revision=revision, local_dir="artifacts/GS1-0.8B")
print(f"Model revision: {revision}")
PY

Save the printed model revision and the framework commit (git rev-parse HEAD) with your results. To repeat a download, set revision to the saved model SHA.

Run a decision locally

Save as quickstart.py and run with uv run python quickstart.py:

from typesafe_sdk import Choice, Noul, Score

from yes1.api.schema import DecisionRequest
from yes1.inference.engine import Engine
from yes1.model.artifact import load_artifact

artifact = load_artifact("artifacts/GS1-0.8B", device="cuda", precision="bf16")
artifact.model.eval()
engine = Engine(
    artifact.model,
    artifact.compiler(),
    mode="reference",
    precision="bf16",
    reference_batching=False,
)

request = DecisionRequest(
    model="GS1",
    state={"battery": 18, "delivery_cost": 12, "required_reserve": 10},
    questions={
        "action": Choice(
            instructions="Choose an action that preserves the required battery reserve.",
            criteria={"deliver": "Deliver now.", "recharge": "Recharge first."},
        ),
        "safe": Noul(
            instructions="Can delivery finish with at least the required battery reserve?",
        ),
        "risk": Score(
            instructions="Rate the risk of delivering now against the reserve requirement.",
            criteria=["No risk.", "Uncertain.", "Reserve would be violated."],
        ),
    },
)

try:
    result = engine.evaluate(request)
    print(result.model_dump_json(indent=2))
finally:
    engine.close()

Put shared facts in state and submit related questions together. Choice returns choice and probabilities; Noul returns noul = $p(\mathrm{true})$; Score returns score and level probabilities. Score levels start at 0.

Serve over HTTP

uv run yes1 serve --model-path artifacts/GS1-0.8B \
  --device cuda --precision bf16 --engine reference --model-name GS1

In another terminal:

curl http://127.0.0.1:8000/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "GS1",
    "state": {"battery": 18, "delivery_cost": 12, "required_reserve": 10},
    "questions": {
      "action": {
        "type": "choice",
        "instructions": "Choose an action that preserves the required battery reserve.",
        "criteria": {"deliver": "Deliver now.", "recharge": "Recharge first."}
      }
    }
  }'

The server also exposes GET /v1/models and supports the TypeSafe Python SDK. Local requests require no API key.

Keep the complete downloaded artifact, including artifact.json and tokenizer/processor assets. Weights and config.json alone are insufficient for serving. Start with reference; shared execution and CUDA graphs require runtime qualification.

Interactive Playground

With the model server running, open another terminal at the repository root. Requires Node.js 22+:

cd playground
npm ci
npm run build
DEFAULT_BACKEND=yes1_local ../scripts/serve_playground.sh

Open localhost:3000. Use Query to try Choice, Noul, and Score and inspect probabilities, or explore model-driven Go, Maze, and Snake. See the Playground guide for remote backends.

💻 Input Format

Pass structured requests as shown above. The compiler handles XML escaping, the chat template, structural tokens, and the final decision token. Its decision payload has this layout:

<state>
Shared facts and observations
</state>
<question>
Choose the next action.
</question>
<option>
deliver: Deliver now.
</option>
<option>
recharge: Recharge first.
</option>

Choice supports 1–255 candidates; Score requires 2–10 ordered levels. Oversized requests are rejected rather than silently truncated.

Use meaningful option keys and precise descriptions: Choice keys enter the model input. To allow abstention, include an explicit option such as unknown: Insufficient evidence to decide.

Training Data and Recipe

Data provenance: The repository provides adapters for public annotated datasets and independently verified synthetic decision tasks. A checkpoint-specific dataset inventory, corpus manifest, and training logs are not included; supported sources should not be read as a confirmed training mixture. See the data guide.

The maintained 0.8B recipe specifies the defaults below, rather than a verified record of this checkpoint's training:

Phase	Objective	Updates	Peak learning rates: head / language / router
Alignment	NLL	100	2e-4 / 4e-5 / 1e-4
Finetuning	NLL + Brier	Corpus-dependent	2e-5 / 1e-5 / 1e-5

Both phases use BF16, a global batch of 4,096 logical questions, a 16,384-token context limit, seed 12, and gradient checkpointing. Alignment trains structural-token offsets in the language group. These supervised defaults do not document the CRPO or multi-step RL experiments described above.

Intended Use and Limitations

Use GS1 for candidate selection, routing, verification, extraction, and ordinal scoring within agent workflows.

• Decisions depend on the supplied evidence and candidate set. Include an explicit abstention option when evidence may be insufficient.
• Probabilities can be miscalibrated under distribution shift. Validate accuracy, calibration, and abstention on your own workload before setting thresholds.
• Biases from the base model and training data may affect decisions; use human review for consequential applications. Reported text benchmarks do not establish multimodal or safety performance.

🛠️ Finetuning and Export

To finetune this checkpoint on an existing verified corpus:

uv sync --locked --extra train --extra sources --extra decision-index --extra dev
uv run python scripts/train_gs1.py --config scripts/configs/gs1_0_8b.yaml \
  --base-model artifacts/GS1-0.8B --corpus /path/to/frozen-corpus \
  --run-dir runs/gs1-0.8b --wandb-mode offline

# Serve the complete export produced after training.
uv run yes1 serve --model-path runs/gs1-0.8b/exports/GS1 \
  --device cuda --precision bf16 --model-name GS1

Native GS1 inputs go directly to finetuning. Raw Qwen3.5 inputs run alignment first. For initialization from Qwen:

from yes1.model import GS1Model
from yes1.model.artifact import export_artifact

model, tokenizer = GS1Model.initialize("/path/to/Qwen3.5-0.8B")
export_artifact(model, tokenizer, "artifacts/GS1-untrained", status="untrained")
model = GS1Model.from_pretrained("artifacts/GS1-untrained")

Qwen initialization still requires decision training. Export to a new, empty directory with export_artifact; save_pretrained alone does not create a complete serving artifact.

See the training guide for multi-GPU recipes and resume, and the deployment guide for multi-GPU serving.

Citation

If you find GS1 useful, please 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}
}