hs-decide-1-flash

hs-decide-1-flash

Answers typed questions about a text state with a probability for every option in one pass, with no generated text, in eight languages. A 0.8B decision model for the Hindsight tool, served with vLLM or Transformers.

Made by hindsight-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: hs-decide-1-flash is a small decision model from hindsight-ai, made for use with Hindsight. It has 0.8B parameters and starts from Qwen3.5-0.8B. You hand it a situation described in text, together with typed questions about the situation. It replies with the likelihood of each possible answer and a confidence figure, all in a single pass, and never produces prose.

Who it suits: Developers who want fast, structured decisions from a small model, such as classifying tickets or choosing between documents, and who use or build on Hindsight.

Industry | Who uses it | How they use it | Example
Customer support | Support engineers | Classify tickets with yes or no questions | Is a ticket a billing problem
Search and retrieval | Retrieval engineers | Choose the best document for a query | Which document gives the current date
Agent building | Agent developers | Score how well a document answers a question | Rating a passage on four levels
Multilingual products | Global teams | Run decisions in eight languages | A decision on a Japanese message

If you want a model that writes text, you can skip it, because it only returns probabilities.

Where to use it:

• Hindsight, which it was made for
• vLLM, through a plugin in the repository's serving folder
• The Transformers library, on a plain CPU, a Mac with Apple chips or a PyTorch-compatible GPU

Use case | Who | What they get back
Yes or no checks | Support teams | The probability that the answer is yes
Choosing among options | Retrieval teams | A probability for each of 2 to 255 options
Ordered ratings | Quality teams | A score on 2 to 10 ordered levels, with probabilities

How to set it up:

• Python and pip, with Python 3.12 or later for the Transformers route
• PyTorch installed for your hardware, for the Transformers route
• About 1.5 GB of disk space in BF16, or about 376 MB at 4-bit (an estimate)

1. Install the Hugging Face hub tools and download the model with its serving folder.
2. Install the serving package.
3. Start a server or call the model from Python.

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

For the Transformers route, fetch the serving folder and install it with its server extra.

hf download hindsight-ai/hs-decide-1-flash --include "serving/*" --local-dir hs-decide-1-flash
pip install './hs-decide-1-flash/serving[server]'

How to use it: Load the model in Python, pass a state and a set of typed questions, and read the probabilities. A choice question has 2 to 255 options, a score question has 2 to 10 ordered levels and a noul question is yes or no.

from hindsight_decision import DecisionModel

model = DecisionModel.from_pretrained("hindsight-ai/hs-decide-1-flash")
response = model.answer({
    "state": "Ticket: I was charged twice for my March invoice.",
    "questions": {"billing": {"type": "noul", "instructions": "Is this a billing problem?"}},
})
print(round(response["answers"]["billing"]["noul"], 3))
# 0.997

To serve it to Hindsight with the Transformers route, start a server. It answers one request at a time.

hindsight-decision-serve-transformers hindsight-ai/hs-decide-1-flash --port 8000 --api-key <key>

Tool | What it does
DecisionModel.from_pretrained | Loads the model in Python
model.answer | Takes a state and typed questions and returns probabilities
hindsight-decision-serve | Starts a vLLM server for the model
hindsight-decision-serve-transformers | Starts a Transformers server, one request at a time

Example: You give it the ticket text about being charged twice for a March invoice, with the question of whether it is a billing problem. The card's run prints about 0.997, the probability of yes.

For newcomers and experts: If you are new to this, run the short Python example first and change the ticket text to see how the probability moves. It needs no server.

The model was trained on requests of up to 65,536 tokens and answers all questions about one state in the same pass. If Hindsight's requests time out, the card suggests lowering the maximum request tokens so the server refuses long requests and Hindsight splits them.

You write a state
You add typed questions
The model scores every option in one pass
You read the probabilities

Good to know:

• Throughput: the Transformers server answers one request at a time
• Languages: eight in all, including English, Japanese, Korean and Chinese
• License: apache-2.0
• Cost to run: no paid service is named, only your own hardware, and the server can take an API key you set
• Data it reaches: downloading fetches model files, and the server reads the requests you send it
• Not for: writing text, since it returns probabilities and generates none

On Hugging Face, 13 November 2026

• Task: text ranking
• Library: transformers
• Parameters: 752M
• Weights: about 1.5 GB in BF16; about 376 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 858
• Likes: 20
• License: apache-2.0
• Languages: en, es, fr, de, pt, ja, ko, zh
• Based on: Qwen/Qwen3.5-0.8B
• Quantized versions: 1
• Fine-tunes: 2
• Tags: transformers, safetensors, qwen3_5_text, text-generation, decision-model, reranker, agent-memory, vllm, hindsight, text-ranking

Links

• Model: https://huggingface.co/hindsight-ai/hs-decide-1-flash
• Files: https://huggingface.co/hindsight-ai/hs-decide-1-flash/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download hindsight-ai/hs-decide-1-flash

Then follow the usage in its model card. Nothing in this kit runs by itself.