d3 is a new Ollama model that Makes DECISIONS (500 Messages in 23 Seconds!) - d3 on a Laptop GPU

Author: Prompt Engineer 48 · Watch on YouTube

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: d3 is a new Ollama model that Makes DECISIONS (500 Messages in 23 Seconds!) - d3 on a Laptop GPU · Channel: Prompt Engineer 48 · 03:06 · Watch on YouTube

Topics: AI evaluation and benchmarks, Model inference and optimization, AI models, Local and open-source models

Summary

The video tests D3, an Ollama model designed to score choices rather than generate conversational text, on ticket routing and screenshot classification. In a 500-message synthetic ticket test, D3 2B achieved 95.8% accuracy at 23 messages per second, roughly matching the QN 34B instruct model’s accuracy while running somewhat faster. D3 also provides confidence scores: the 406 messages it rated at least 90% confident were all classified correctly in this test. In a separate test with 30 synthetic screenshots, both models got every example right, while D3 2B finished much faster. The presenter recommends D3 for routing, triage, and yes/no checks, but a chat model for explanations or free-text responses.

Key points

D3 returns scores and probabilities for options supplied by the user instead of generating a conversational answer.
The video says D3 is available through Ollama in sizes of 2B, 4B, 9B, and 27B; the 2B package is about 3 GB.
The presenter says Ollama 0.40.9 or newer is required and uses version 40.11.
In a 500-message synthetic ticket-routing test, D3 2B scored 95.8% accuracy and processed 23 messages per second.
QN 34B instruct scored 95.6% accuracy and processed 19 messages per second in the comparison.
D3 confidence scores allowed the presenter to select 406 messages rated at least 90% confidence; all 406 were correct, while the other 94 were 78% correct.
The presenter cautions that the ticket messages were generated from templates and the screenshots were clean and synthetic.
In a 30-image classification test, both models got all examples right; D3 2B took 8.6 seconds versus 58 seconds for Qwen 3VL.
The video recommends D3 for routing, triage, or yes/no checks, and chat models for explanations or free text.

Chapters

Start	Chapter	Summary
00:00	The routing problem and D3’s approach	The presenter introduces a ticket-sorting task, shows a chat model failing to return a category, and explains how D3 scores supplied choices instead.
00:30	Running D3 and comparing ticket routing	The video demonstrates D3 through Ollama and its API, then describes the 500-message test setup and reports the models’ accuracy and speed.
01:31	Results and confidence-based routing	The presenter compares the ticket-routing results, examines D3’s confidence scores, and notes the limitations of the synthetic test data.
02:34	Screenshot test and verdict	Both models classify 30 synthetic screenshots correctly, after which the presenter summarizes where D3 is useful and when to use a chat model.

Quotes

And the model that did it never wrote a single word. — 00:00

It scored your options and that one detail is why it’s so fast. — 01:01

So, on accuracy, it’s a tie and I’m not going to pretend otherwise. — 02:02

These are clean, synthetic screens, so this shows it can do the job, not that it’s perfect on messy, real apps. — 02:34

If your job is routing, triage, or yes/no checks, D32B is a great fit. — 03:06

Concepts and entities

D3, Ollama, vLLM Semantic Router, Qwen, Qwen3-VL, RunPod, Pontuação de confiança, Roteamento de mensagens

Full transcript (original language)

00:00 500 customer messages, 23 seconds, 95.8% right on a laptop with 8 gigs of video memory. And the model that did it never wrote a single word. I’m going to show every step on my own screen and I’m going to race it against a normal QN model because that’s the only way to know if this is real. >> [music] >> Here’s the problem today. You want to sort support tickets. So you ask a chat model for one word. Look at what I got from the stock QN 34B on my very first try. It started reasoning out loud, ran

00:30 out of tokens, and never gave me a category. It also spilled onto the CPU because its default context window is huge. So you end up writing parsers, retries, and guardrails. D3 is built to skip all of that. D3 comes from the VLLM semantic router project and it’s on Ollama now. The page says it right here. You give it text, structured data, or an image and it returns scores and probabilities for the options you provide. It does not write a conversational answer. There are four sizes, 2, 4, 9, and 27 billion. The 2B package is about 3 GB, so it fits my card with room to spare. You need Ollama

01:01 0.40.9 or newer. I’m on 40.11. Let’s run it. You ask a question, you give it choices, and that’s all. Billing 99%. A cracked screen goes to returns. An app crash goes to technical. Now the API. Same idea, but look at the response. A probability for every choice, one confidence number, and output tokens zero. It never generated text. It scored your options and that one detail is why it’s so fast. The test. 500 messages, five teams, billing, returns, technical, shipping, and account. I know the right answer for

01:31 every message. The same five category descriptions go to both models. The QN is the 34B instruct version, which doesn’t think out loud with its context capped at 4,000 tokens, so it stays on the GPU. One honest caveat, I generated these messages from templates, so this set is easier than real tickets. Keep that in mind. Results. Clean runs, nothing else open. D3 2B, 95.8% 23 messages a second, 40 milliseconds each. QN 34B instruct, 95.6% 19 messages a second, 49 milliseconds.

02:02 So, on accuracy, it’s a tie and I’m not going to pretend otherwise. D3 is a bit faster and both gave identical answers on the second run. The D34B reaches 98.4% but it’s slower, six messages a second because it’s the bigger package. Here’s what Qwen can give you. Every D3 answer comes with a confidence. Keep only the answers where it’s 90% sure and that’s 406 messages. All 406 were right. The other 94, the unsure ones, were only 78% right. So, you can auto route the confident ones and send the rest to a human. The page says, “These scores are model outputs, not guarantees, so test

02:34 it on your own data.” Second experiment, 30 screenshots I generated, 10 error screens, 10 login screens, 10 success screens. Both models got 30 out of 30. The difference is time. D32B finished in 8.6 seconds. Qwen 3VL took 58 because it reasons first about 70 tokens before it answers. These are clean, synthetic screens, so this shows it can do the job, not that it’s perfect on messy, real apps. So, verdict. If your job is routing, triage, or yes/no checks, D32B is a

03:06 great fit. Small, fast, no parsing, and a confidence score you can act on. If you need explanations or free text, use a chat model. If you want more accuracy on harder decisions, D3 goes up to 27 billion. If you’re looking for cheap and reliable GPUs, I use RunPod and the link is in the description. Tell me in the comments, what would you route first with a model like this?