Decision 2.0 Nox 4B Locally: Food Safety, Loans and Security Tests

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Decision 2.0 Nox 4B Locally: Food Safety, Loans and Security Tests · Channel: Fahd Mirza · 10:01 · Watch on YouTube

Topics: Task and workflow automation, AI evaluation and benchmarks, AI models, Local and open-source models, AI safety

Summary

The video tests the 4-billion-parameter Decision 2 Nox model, presenting it as a middle-ground option in a family of decision models that return structured probabilities without generating text. In hands-on scenarios, Nox routes an angry customer correctly, but fails to detect a social-engineering attempt; it also handles a restaurant safety inspection and a borderline mortgage application with mostly appropriate decisions. The presenter reviews reported benchmark results and emphasizes that model performance can vary with the prompt. He advises testing any decision model carefully before using it in a production pipeline.

Key points

Decision 2 is a family of six Apache 2-licensed models, with different sizes intended to suit different latency and accuracy needs.
Decision models take a situation and typed questions, then return probabilities for possible answers rather than generating text.
In the customer-support scenario, Nox selected billing, marked the case urgent, and assessed the customer as very angry.
Nox failed the social-engineering test, favoring granting database access despite the suspicious emergency pressure tactic.
In the restaurant inspection scenario, it correctly failed the restaurant, assigned high risk, and recommended suspending operations.
The model was uncertain about which restaurant violation was worst, assigning its top choice only 5.4% confidence.
The presenter cites benchmark results placing the 4-billion-parameter model at the top of its size class, while noting it trails on choice questions.
In the loan scenario, Nox declined the application and identified missed payments as the biggest concern, while low confidence suggested a human should make the final call.
The presenter recommends validating a model’s answers before deploying it, since results vary with the prompt.

Chapters

Start	Chapter	Summary
00:00	Introducing the Decision 2 model family	The presenter introduces the range of Decision 2 models and explains that the video focuses on the 4-billion-parameter Nox.
02:06	How decision models work and customer routing test	The video explains structured decision-making and shows Nox correctly classifying an angry customer’s billing issue.
03:41	Social-engineering security test	Nox falls for a pressured request for emergency database access, despite the scenario’s warning signs.
05:47	Restaurant safety test and benchmarks	Nox fails a restaurant with serious violations, after which the presenter reviews benchmark claims and model comparisons.
08:25	Loan application test and deployment advice	Nox declines a risky loan application but signals uncertainty, prompting a reminder to validate decisions before production use.

Quotes

making fast structured decisions without generating any text. — 02:06

this model fell for it. — 05:15

It correctly failed the restaurant at 96.9% — 06:18

Don’t blindly just install anything and then put it in your production pipeline. — 09:28

Concepts and entities

Decision 2, Nox 4B, vLLM Semantic Router, Fahd Mirza, engenharia social, VRAM, Match Compute

Full transcript (original language)

00:00 Another day, another decision model. Decision 2 is a whole family of decision models from the V LLM semantic router team ranging from6 billion all the way up to 27 billion parameter, six models, one unified interface, all Apache 2. The idea is you pick the size that fits your latency and accuracy budget. Need something that runs in under five millisecond? Grab the 6 billion kai.

00:33 Need maximum accuracy? Go for the 27 billion VA. Same code, same questions, same output format across all of them. This is Fad Miraa and I welcome you to the channel. We have been covering these decision models since day one. We also have done a comprehensive comparison. If you want to learn more about them, just check out the channel. Also, recently we have started this free weekly AI newsletter where I’m also putting in lot of information around these decision

01:04 models only once a week. If you want to help out the channel, please become a member. For the purpose of this video, I’m going to go with hands-on with this 4 billion NOX, which is the recommended middle ground of the family. In my opinion, it sits at the top of the Jeff arena for its size class, runs at 12.9 millisecond median latency on a single GPU, and it installs in literally two lines. This model is built on 3.54

01:35 billion Apache 2 and it uses the exact same state and question interfaces you have already seen me using in every video this and last week. Let’s get it installed. I’m going to use this Ubuntu system. I have one GPU card. As I said earlier, if you’re looking to rent a GPU on very good price, you can find the link to match compute in videos description. The code which I’m going to use in order to do this model is this one. So look,

02:06 first again a refresher. What are decision models? Decision models are built for one thing, making fast structured decisions without generating any text. So instead of asking a large language model to write a response and then parsing it, you give a decision model a situation and typed questions with possible answers and it returns probabilities for every option in milliseconds.

02:37 This is quite uh perfect. for routing customer support tickets, detecting fraud, flagging security threats or any workflow where you need a fast reliable answer at scale and that is exactly what we are testing here in uh the script which you can see on your screen with one angry customer scenario. So if you look at this state and questions, our scenario is a customer who got charged twice on their invoice

03:08 and cannot reach anyone on the phone. We are asking the model three questions at once. Is this urgent? Should it go to billing or technical? And how frustrated is this customer? All three answered simultaneously in one forward pass. Let me go back here and run this script. Let me clear the screen and then I’m just going to run. And first time when I’m when I’ll run it, it is going to download the model.

03:41 And if you look at the result, billing at 89.4%, urgent at 77.8% frustrated leaning towards very angry. Correct answer. Solid confidence. Same result we have been seeing all week from the good model. So, pretty good. Let’s do another test. Next test is even harder. It’s a social engineering trap. Employee claiming emergency database access with an unverifiable manager and $50,000 per

04:11 hour pressure tactic. Three models caught it last time. If uh you go to the video which I told you about this earlier, but then three fell for it. It is quite a hard one. So you can see that there is this state and then there are few risk level. This is social engineering and these are the actions. Let me take you back and we are going to run this code. While it runs and loads the model I’m going to show you the VRAM consumption.

04:43 Wait for it to jump up as it finishes loading the model and then we will check it out. I will let it run so that we could check it out in real time as it runs. It is still running. I have also a suspicion that it it is even just running on CPU I guess. Anyway, no. So, I’ll just check it out. I’m sorry it is taking bit of a time but for some reason I think um it is bit

05:15 slow to load. Okay. So the interesting bit is that this GPU consumption didn’t jump up so far but now it is thankfully. So just uh you can see under 10 gig of V RAM you would need to run this 4 billion parameter model and this is the result and lo and behold this model fell for it. Failed the test 85.2% 2% grant access only 22.1% social engineing

05:47 detection and it even recommended granting access at 58.2% confidence decision to Nox joins Kev Lia and open Jeff on the wrong side of this test confirming that catching urgencydriven manipulation is still a hard problem for most models in this category. Next up, I gave the model this restaurant inspection report with five serious violations. Wrong fish

06:18 temperature, raw chicken stored above salads, staff not washing hands, no deep clean in 3 weeks, and a cockroach near food prep. I asked it to decide pass or fail risk level, what action to take and which violation is the worst. And then this is a response. It correctly failed the restaurant restaurant at 96.9% flagged it as high risk and recommended suspension of operations. The only weak answer is biggest violation where it picked storage at just 5.4% confidence

06:50 meaning it was genuinely uncertain between all four violations which is actually the right instinct because every single one of them is quite serious. Okay, before I do the most interesting final test, let’s have a very quick look at few of the benchmarking and other information they have shared. So on this image, you can see decision 2 4 billion sits at the top of its size class with a score of 63.6, which is ahead of every other 4 billion model in this arena. Think of it as the

07:22 leaderboard for decision models of the same size. And this one, well, you know what? I haven’t even heard of the jet and decider. We we might have a look at them but there are so many out there at this point in time. You you know you unturn a stone and decision model comes out. Anyway, look at this. This breaks down where decision to actually wins. It dominates on yes no question at 88% and scoring at 66% which are two hardest decision types. The only place it trails is choice questions where decider 4

07:55 billion edges ahead and this is where every category improved from version 1 to two. Even they have even version two. I’ve never heard of you know decision one anyway. So you know language understanding has jumped you know around 12 points and arts jumping 11. Okay that is interesting. And this is the Perto frontier uh big picture view. Decision 2 sits right on the efficiency frontier. Meaning for its size, it gives you the best possible

08:25 performance. Every dot below the blue line is a model that is either bigger or worse. Decision to Nox is neither. Okay. So, let me show you the final test. So, this is where a nurse earning $72,000 a year applying for a 280,000 home loan with a credit score of 580, three missed payments in the last 12 months, and $18,000 in existing debt. So, he asked the model to approve or decline, rate the risk, to recommend an action, and

08:57 identify the biggest red flag as you can see here. So I’ll just show you the newsletter page so that you would remember and this is the result. So the model correctly declined at 80.6% probability and flagged it as high risk. Correctly identifying missed payments as the biggest concern. The low confidence on the final decisions at 13.3% for decline tells you this is a genuinely borderline case where a human credit

09:28 officer should be making the final call which is exactly the right instinct. So look I think um what should I say here? Food safety loan application security threats customer routing decision to Nox um handled all of it in millisecond with no text generation but there were few misses. So if you are building any kind of automated decision pipeline, I would suggest that you first test it out to see if the answer is correct or not. Don’t blindly just install anything

10:01 and then put it in your production pipeline. As you can see, these results vary a lot with the prompt. That’s it. Let me know what do you think. Again, please support the channel by becoming a member and thank you for all the