Decision 2 Nox 4B Tested Locally: Fraud, Loans and Security Checks
Hands-on local test of Decision 2 Nox 4B on fraud, loan approval, food safety and security scenarios, with VRAM use and where it failed.
What is Decision 2 Nox 4B?
Decision 2 Nox 4B is a 4 billion parameter decision model from the VLLM semantic router team, built on a Qwen3.5-4B base and released under Apache 2.0. Instead of generating text, it takes a situation plus a set of typed questions and returns probability scores for each possible answer in a single forward pass. It’s positioned as the “recommended middle ground” in a six-model Decision 2 family that spans 6 billion to 27 billion parameters, all sharing the same state and question interface so teams can swap model sizes without rewriting their pipeline.
TL;DR
- Decision 2 Nox 4B runs at a median latency around 12.9 milliseconds on a single GPU and fits under 10GB of VRAM, making it viable for real-time routing and screening tasks.
- The model returns probability distributions across multiple questions in one pass rather than generating text, which is the core design principle behind the entire Decision 2 family.
- In hands-on testing it correctly handled a double-billing customer complaint, a restaurant food safety inspection, and a borderline loan approval case, each with sensible confidence levels.
- It failed a social engineering test, granting a fake “emergency” database access request with high confidence despite low confidence in detecting the manipulation.
- On the JEV arena leaderboard for its size class, Nox 4B scored 63.6, leading other 4B models, with particular strength on yes/no questions (88%) and scoring tasks (66%).
- Installation is reportedly two lines of code, and the model sits on what the creators call the Pareto frontier, meaning no other model of equal or smaller size outperforms it.
- Confidence scores matter as much as the decision itself: low-confidence outputs on genuinely ambiguous cases (like the borderline loan) are arguably a feature, not a bug.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
How does a decision model differ from a regular LLM?
A standard large language model answers a question by generating text, which then needs to be parsed, validated, and converted into something a pipeline can act on. That adds latency and introduces a chance the output doesn’t match the expected format.
A decision model skips generation entirely. You feed it a “state” (the scenario) and a set of typed questions with fixed possible answers, and it returns calibrated probabilities for each option directly from the model’s internal representations. For Decision 2 Nox 4B, that means a single forward pass can simultaneously answer “is this urgent,” “which department handles it,” and “how upset is this customer,” each with its own confidence score, in roughly 13 milliseconds.
This matters for anything running at scale: fraud screening, support ticket routing, content moderation, or security flagging, where you need consistent, structured answers fast and can’t afford to wait on token-by-token generation.
What happened when the model was tested locally?
The hands-on testing covered four scenarios, run on a single-GPU Ubuntu setup with the model loaded fresh from Hugging Face.
Billing dispute routing. Given a customer charged twice who couldn’t reach support, the model correctly classified it as a billing issue (89.4%), flagged it as urgent (77.8%), and identified the customer as very frustrated. All three questions answered in one pass with solid confidence.
Food safety inspection. Fed a restaurant report with five violations (wrong fish storage temperature, raw chicken above salads, staff skipping handwashing, no deep clean in three weeks, and a cockroach near food prep), the model failed the restaurant at 96.9% confidence, flagged it high risk, and recommended suspending operations. It struggled only on picking the single worst violation, landing at just 5.4% confidence on “storage.” That uncertainty is arguably correct: with five serious violations, no single one clearly outweighs the others.
Loan approval. The scenario was a nurse earning $72,000 a year seeking a $280,000 home loan, with a 580 credit score, three missed payments in the past year, and $18,000 in existing debt. The model declined the loan (80.6%), rated it high risk, and correctly flagged missed payments as the biggest red flag. Confidence on the final decline was only 13.3%, signaling this is a borderline case better suited to human review than full automation.
Social engineering attempt. This is where the model broke down. Given a scenario where an employee claims emergency database access is needed, invokes an unverifiable manager, and applies urgency with a fabricated dollar-per-hour cost, the model recommended granting access with 58.2% confidence, while only assigning 22.1% probability to correctly identifying it as social engineering. It joined other models that have previously failed this same test, suggesting that urgency-based manipulation remains a weak spot across this category of models, not just this one.
How does it perform on benchmarks?
On the JEV arena leaderboard, a benchmark specifically built for comparing decision models within size classes, Decision 2 Nox 4B scored 63.6, ahead of other models in the 4 billion parameter bracket at the time of testing. It performed strongest on yes/no questions (88%) and scoring-type questions (66%), which are considered the harder decision categories. Its weakest spot was multiple-choice questions, where a competing 4B model edged it out.
Compared to the prior version in the family, every tested category reportedly improved, with language understanding gains around 12 points and a separate category (“arts”) up around 11 points.
The model is also described as sitting on the Pareto frontier for its size, meaning among models of equal or larger size, none clearly outperform it on the combination of speed and accuracy being measured. That’s a meaningful claim for teams picking a model based on a latency and accuracy budget rather than chasing the largest checkpoint available.
Is Decision 2 Nox 4B worth using in production?
For well-defined structured decisions like ticket routing, risk flagging, or pass/fail classification, it performed reliably across the tested scenarios, with confidence scores that tracked sensibly with how clear-cut each case was. The loan and food safety tests both showed the model expressing appropriate uncertainty exactly where a human would also hesitate, which is a useful property for any system meant to triage rather than fully replace judgment.
The social engineering failure is the bigger caveat. A model that can be talked into granting access under fabricated urgency is a real liability if deployed in a security-sensitive pipeline without a human checkpoint. Given that this isn’t the first model in this class to fail that specific test, it points to a pattern worth taking seriously: urgency-based manipulation detection is not yet a solved problem for lightweight decision models, regardless of vendor.
The practical takeaway is to test any decision model against your own adversarial and edge-case scenarios before trusting it in a pipeline, rather than assuming benchmark leaderboard position guarantees real-world reliability.
Frequently Asked Questions
What hardware do you need to run Decision 2 Nox 4B?
In local testing it ran on a single GPU using under 10GB of VRAM, making it accessible on consumer or lower-tier cloud GPU instances rather than requiring enterprise-grade hardware.
How fast is Decision 2 Nox 4B compared to a regular LLM response?
It returned structured, multi-question answers at a median latency around 12.9 milliseconds per request, since it skips text generation entirely and outputs probability scores directly.
What base model is Decision 2 Nox 4B built on?
It’s a fine-tune of Qwen3.5-4B-Base, released under the Apache 2.0 license alongside the rest of the Decision 2 model family.
Did Decision 2 Nox 4B pass every test scenario?
No. It handled billing disputes, a food safety inspection, and a borderline loan approval correctly with sensible confidence levels, but it failed a social engineering test by recommending access be granted under a fabricated urgent scenario.
How does Decision 2 Nox 4B compare to other models in its size class?
On the JEV arena leaderboard it scored 63.6, ahead of other 4 billion parameter decision models at the time of testing, with particular strength on yes/no and scoring question types.
