Torchcast Decision 4B

Torchcast Decision 4B

A 4B decision model built on Qwen3.5-4B that returns yes/no probabilities, choice distributions and score distributions in one pass without generating text. For English classification, routing and rubric checks.

Made by torchcast-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: Torchcast Decision 4B is a decision model from Torchcast AI, built on Qwen3.5-4B. It returns probabilities for yes or no questions, distributions over choices and distributions over ordinal scores. Each decision reads the answer options in one forward pass, and nothing is generated. This is version 1.3.1.

Who it suits: Developers who need a fast classifier that outputs probabilities, such as for classification, routing and bounded rubric evaluation of English text.

Industry | Who uses it | How they use it | Example
AI engineering | Teams building routers | Pick a route from a list of options | Choose which queue gets a message
Evaluation | People scoring model outputs | Grade an answer against a rubric | Return a score distribution for a reply
Operations | Teams that classify text | Answer yes or no questions about a text | Flag whether a ticket needs escalation

You can skip it if you need it to act, since the server returns decisions and does not run actions. It is also tuned for English text only.

Where to use it:

• The bundled torchcast_decision serving package, with its decision_config.json
• A GPU server, since the card quotes timings on one A40
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Yes or no decisions | Developers | A probability for yes and for no
Choice among options | Router builders | A distribution over the options
Score questions | Evaluators | A distribution over an ordinal scale

How to set it up: Use the bundled serving package with its config file. The card warns that generic text-generation examples and causal-attention runtimes such as vLLM, llama.cpp and MLX do not reproduce the evaluated behavior. Its serving instructions live in the project's code repository.

• Python and pip
• A GPU for the bundled server
• The torchcast_decision package and decision_config.json from this repository

1. Download the repository files.
2. Follow the serving instructions in the project's code repository.
3. Start the bundled server with the bundled settings.

pip install -U huggingface_hub
hf download torchcast-ai/torchcast-decision-4b

How to use it: You send a question and its options to the served model and read the probabilities that come back. Prompts over 16,384 tokens keep their head and tail. The server returns JSON errors, with a 503 status when the GPU runs out of memory.

Example: You send the question "Does this ticket need a human?" with the options yes and no. The server returns a probability for each, and your code takes the higher one.

For newcomers and experts: If you are new, think of it as a judge that gives odds instead of sentences. Start with one yes or no question and read both probabilities before building routing on top.

For experts, the full-attention layers read the whole prompt in both directions, so loading the weights in plain transformers gives different probabilities. Version 1.3 is a new full fine-tune of Qwen3.5-4B. By using CUDA graphs, version 1.3 roughly halves the median time per request compared with 1.2 on one A40.

Question and options
One forward pass, no generation
Probabilities for each option

Good to know:

Title: JevBench composite score, as listed in the card
Torchcast Decision 4B v1.3: 76.16
H2O-Lightning-4B v1.1: 73.78
Torchcast Decision 4B v1.2: 73.00

• Results: the figures were measured by the leaderboard's maintainers, and the card says scores may be re-placed
• Limits: probabilities can be unreliable on new domains
• Language: English
• License: Apache 2.0, while an earlier release keeps a non-commercial license
• Cost to run: no API key or paid service named, but it needs a GPU
• Data it reaches: downloads files from Hugging Face, nothing else is named
• Not for: taking actions, other languages, or loading in generic runtimes

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 5.2B
• Weights: about 10.3 GB in BF16; about 2.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,434
• Likes: 10
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 1
• Fine-tunes: 4
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, typed-decisions, jevbench, conversational

Links

• Model: https://huggingface.co/torchcast-ai/torchcast-decision-4b
• Files: https://huggingface.co/torchcast-ai/torchcast-decision-4b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download torchcast-ai/torchcast-decision-4b

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