hs-decide-1-flash

hs-decide-1-flash is a 0.8B decision model for Hindsight. You give it a state and one or more typed questions about it, and it returns a probability for every answer option in a single forward pass, with no generated text.

Highlights

• Typed questions: choice (pick one of 2 to 255 options), score (rate on 2 to 10 ordered levels) and noul (yes or no).
• Probabilities: a probability for every option, plus a confidence.
• Many questions per request: all questions about one state are answered in the same pass.
• Multilingual: English, Spanish, French, German, Portuguese, Japanese, Korean and Chinese.
• Long requests: trained on requests up to 65,536 tokens.
• Serves on vLLM with the endpoint plugin in this repository's serving/ folder.

Model overview

Model type	Decision model
Base model	Qwen/Qwen3.5-0.8B
Parameters	0.8B
Languages	en, es, fr, de, pt, ja, ko, zh
Context length	65,536 tokens per request (trained)
Question types	choice, score, noul
Precision	bfloat16

Evaluation

In Hindsight

Against other 0.8B decision models

Usage

vLLM

The serving/ folder is a vLLM plugin that adds a POST /v1/systemone endpoint to vllm serve.

hf download hindsight-ai/hs-decide-1-flash --local-dir hs-decide-1-flash
pip install vllm==0.32.0 ./hs-decide-1-flash/serving
hindsight-decision-serve hs-decide-1-flash --port 8000

Flags after the model folder go to vllm serve, for example --api-key or --gpu-memory-utilization.

Request

curl 127.0.0.1:8000/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "hs-decide-1-flash",
    "state": {
      "query": "When is the team offsite?",
      "documents": {
        "D00": "The offsite moved from April 10 to April 17.",
        "D01": "The team offsite is on April 10 in Denver."
      }
    },
    "questions": {
      "latest": {
        "type": "choice",
        "instructions": "Which document gives the current offsite date?",
        "criteria": {"D00": null, "D01": null}
      },
      "moved": {
        "type": "noul",
        "instructions": "Did the offsite date change?"
      },
      "D01": {
        "type": "score",
        "instructions": "How well does document `documents.D01` answer `query`?",
        "criteria": ["about something else", "same topic, no help", "helps", "answers"]
      }
    }
  }'

Response

{
  "model": "hs-decide-1-flash",
  "answers": {
    "latest": {
      "type": "choice",
      "choice": "D00",
      "confidence": 0.91,
      "probabilities": {"D00": 0.95, "D01": 0.05}
    },
    "moved": {"type": "noul", "noul": 0.97},
    "D01": {
      "type": "score",
      "score": 1.4,
      "confidence": 0.62,
      "legend": {"0": "about something else", "1": "same topic, no help", "2": "helps", "3": "answers"},
      "probabilities": {"0": 0.03, "1": 0.55, "2": 0.38, "3": 0.04}
    }
  }
}

A request longer than 65,536 tokens returns 400 with error_type: max_tokens_exceeded.

Hindsight

Start the server with --api-key, then point Hindsight's decision model at it:

HINDSIGHT_API_DECISION_MODEL_PROVIDER=hindsight
HINDSIGHT_API_DECISION_MODEL_BASE_URL=http://localhost:8000
HINDSIGHT_API_DECISION_MODEL_API_KEY=<the server's --api-key>

Hindsight Inference

hs-decide-1-flash is also served at https://inference.hindsight.vectorize.io. To use it from Hindsight:

HINDSIGHT_API_DECISION_MODEL_PROVIDER=hindsight
HINDSIGHT_API_DECISION_MODEL_API_KEY=<your Hindsight Inference API key>

License

Apache 2.0.

Citation

@misc{hindsightai2026hsdecide1flash,
  title  = {hs-decide-1-flash},
  author = {{Hindsight AI}},
  year   = {2026},
  url    = {https://huggingface.co/hindsight-ai/hs-decide-1-flash}
}