Seb-9B-v2

Seb-9B-v2

A 9B decision model based on Qwen3.5-9B. You give it a situation, a question and options, and it returns how likely each option is, in one pass with no text generated.

Made by ironbcc 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: Seb-9B-v2 is a compact model that decides things rather than writing prose. Feed it a scenario written as text or JSON, add a question and several answer options, and it returns a likelihood for each one.

It does this in one forward pass, so nothing is generated token by token. The card gives ticket routing, phishing detection and choosing a robot's next action as examples of the same call.

Who it suits: Developers who need fast classification or routing inside an app, such as sorting tickets, flagging risky messages or picking a next step. It is built on Qwen3.5-9B.

Industry | Who uses it | How they use it | Example
Customer service | Support platform teams | Route tickets to the right team | Choosing a team from ten options
Security | Email and abuse teams | Flag phishing and jailbreak attempts | Scoring a message as phishing or not
Robotics | Robot developers | Pick the next high-level action | Choosing which object to sort next
AI platforms | LLM gateway builders | Route requests to a suitable model | A first-tier router in front of larger models

You can skip it if you need a model that writes answers or essays, or one with deep general knowledge and reasoning.

Where to use it: It runs from Python with the transformers library, and the card describes use on devices and servers.

• Application back ends that need a quick decision
• Edge devices, since the card reports timings on an Apple M5 Max laptop chip
• GPU servers, with timings on an RTX PRO 6000
• Robot and game control loops, as the card's demos show

Use case | Who | What they get back
Ticket routing | Support teams | A probability for each team option
Message screening | Security teams | A likelihood that a message is phishing
Robot decisions | Robot developers | A chosen next action with option probabilities

How to set it up: You need Python and enough memory for a 9B model, about 19.3 GB in BF16. The visible card gives only the download steps, and it points to a blog post for edge use, routing and robot demos.

• Python with pip
• About 19.3 GB for BF16 weights, or about 4.8 GB at 4-bit as an estimate
• A GPU or a recent Apple chip for the speeds the card reports

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Load the model with your usual transformers workflow and pass it a situation, a question and options.

pip install -U huggingface_hub
hf download ironbcc/Seb-9B-v2

How to use it: Frame each decision as a situation, a question and a set of options. The card describes three question types, and the model answers with a probability for every option.

Type | Options | Example
noul | Y or N | Does the review mention a refund?
choice | Letters A to T, up to 20 options | Which team should handle this ticket?
score | Numbers 0 to 4 | How urgent is the message?

Example: You send a customer message plus the question of which team should handle it, with options A to D. The model returns a probability for each team in one pass, and you act on the highest one.

For newcomers and experts: If you are new, think of it as a multiple-choice grader rather than a chatbot. Start with a yes or no question about a short text and read the probabilities it returns.

If you know the field, it is a 9B classifier head over Qwen3.5-9B that returns option probabilities from a single forward pass. Per decision, the card lists around 22 ms with an RTX PRO 6000 GPU and around 88 ms with an Apple M5 Max. Calibration is one step back, since v1 is still better calibrated.

Situation, question and options go in
One forward pass runs
Probabilities come out per option
Your app takes the top choice

Good to know: The card reports a mean accuracy of 0.801 across 17 public test suites and gains over the first Seb-9B on several tests. It says knowledge and reasoning benchmarks are where a rival model is clearly better.

Title: Mean accuracy across 17 public suites
Seb-9B-v2: 0.801 accuracy
Seb-9B: 0.792 accuracy
Jev 1.13 hosted: 0.786 accuracy
Laya: 0.694 accuracy

Title: Biggest gains over the first Seb-9B
Test | Seb-9B | Seb-9B-v2
BANKING77 | 0.907 | 0.973
Jailbreak detection | 0.868 | 0.940
Model routing | 0.945 | 0.980
Phishing | 0.863 | 0.895
MASSIVE scenario | 0.780 | 0.833

• License: Apache 2.0
• Cost to run: none named, but a 9B model needs a capable GPU or chip
• Data it reaches: the situation text or JSON you feed it, and the files you download
• Not for: writing text, and knowledge-heavy exams where the card says a rival is clearly better

On Hugging Face, 13 November 2026

• Task: text classification
• Library: transformers
• Parameters: 9.7B
• Weights: about 19.3 GB in BF16; about 4.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 265
• Likes: 4
• License: apache-2.0
• Based on: Qwen/Qwen3.5-9B
• Quantized versions: 2
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, decision-model, typed-decisions, classification, routing, calibrated-probabilities, text-classification

Links

• Model: https://huggingface.co/ironbcc/Seb-9B-v2
• Files: https://huggingface.co/ironbcc/Seb-9B-v2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ironbcc/Seb-9B-v2

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