Decision-1.0-Nox-4B

Nox, Latin for night.

Give Nox a state, questions and possible answers. It returns typed decisions and probabilities, with labels defined at runtime.

Decision family

Measured capability

73.09% weighted accuracy across 3,766 decisions and 54 tasks: +2.99 points over Kev-4B on the same benchmark.

Accuracy (%). Overall weights: Decisions 30%, Composition 25%, Reading 15%, Inference 15%, Transfer 15%. These outcome-informed product-priority weights were chosen after observing results; reweighting is not a training improvement. Bold marks Decision-family cells above every external open or untuned reference for that metric, excluding Jev and the other Decision models.

All 54 tasks · Order, missing-evidence and calibration diagnostics · Methods and uncertainty

More questions, one request

Distinct Choice questions at a fixed 499 input tokens per question. Thirty measurements per point across six independently loaded processes on an otherwise idle AMD gfx942 GPU. Python latency includes tokenization and inference; loading and network are excluded. These measurements precede null-description normalization and use explicit descriptions. p50, p95 and measurement scope.

Download the complete model repository

hf download vllm-sr/Decision-1.0-Nox-4B --local-dir Decision-1.0-Nox-4B

This downloads the complete model release. The root config.json lists the backbone, tokenizer, decision head, and calibration files.

Use with 🤗 Transformers

The repository includes its inference code, so stock Transformers can download and run the complete model locally with trust_remote_code=True. system_one takes and returns the same System One request and response bodies as the Decision runtime; nothing is generated.

pip install "transformers>=5.17" torch safetensors huggingface_hub

from transformers import AutoModel

model = AutoModel.from_pretrained("vllm-sr/Decision-1.0-Nox-4B", trust_remote_code=True)
response = model.system_one(
    state="Customer requests a refund.",
    questions={
        "route": {
            "type": "choice",
            "instructions": "Which team should handle this?",
            "criteria": {"billing": "Payments and refunds", "technical": "Product faults"},
        },
        "priority": {
            "type": "score",
            "instructions": "How urgent is this request?",
            "criteria": ["Not urgent", "Soon", "Today", "Immediately"],
        },
    },
)
print(response["answers"]["route"]["choice"], response["answers"]["priority"]["score"])

pipeline("decision", model="vllm-sr/Decision-1.0-Nox-4B", trust_remote_code=True) accepts the same request body. A malformed question is answered with an invalid_question error. The model loads on the first GPU when one is visible, otherwise on the CPU (pass device="cpu" or device="cuda:0" to choose). On a GPU the backbone runs in BF16 with an FP32 decision head; on the CPU everything runs in FP32. A complete question, its candidates and the state are limited to 16,384 tokens; if a question is longer, every question of the request is answered with a max_length_exceeded error and nothing is truncated. flash-linear-attention speeds up the linear-attention layers on a GPU.

Serve with vLLM Semantic Router

This repository contains model data and its Transformers loading code. Use the vLLM Semantic Router Decision runtime to load vllm-sr/Decision-1.0-Nox-4B and serve Choice, Noul, and Score requests. The serving implementation and its dependencies live in vLLM Semantic Router; this release bundles no serving code. The weights do not depend on a particular accelerator; which hardware can serve them is decided by the runtime. For local inference without a server, see Use with 🤗 Transformers above.

After configuring a compatible Decision endpoint, send a SystemOne request (replace the placeholder URL and key):

curl -X POST 'https://your-decision-endpoint.example/v1/systemone' \
  -H 'Authorization: Bearer YOUR_ENDPOINT_API_KEY' \
  -H 'Content-Type: application/json' \
  --data-raw '{"model":"Decision-1.0-Nox-4B","state":"Customer requests a refund.","questions":{"route":{"type":"choice","instructions":"Which team should handle this?","criteria":{"billing":"Payments and refunds","technical":"Product faults"}}}}'

The Hugging Face repository is a model download, not a hosted inference endpoint.

The published model's complete state, question, and candidates have a 16,384-token input limit. See the evaluation scope for measured conditions.

Architecture

A causal Qwen3.5 text backbone combines gated linear and full attention. A shared candidate head reads candidate endpoints and the final query vector. The serving runtime schedules questions according to available hardware and request load.

Candidate head · Vector architecture

Adapted from Qwen3.5-4B. It evaluates supplied evidence without live retrieval; confidence does not guarantee correctness. License · Attributions.