Diffusion at Scale: Testing Mercury 3 by Inception

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: Diffusion at Scale: Testing Mercury 3 by Inception · Channel: Fahd Mirza · 09:57 · Watch on YouTube

Topics: AI evaluation and benchmarks, Software engineering with AI, AI models, Multimodal models

Summary

The video tests Inception’s Mercury 3, a diffusion-based language model, by asking it to fix a subtle bug in a donor-family registry demo and complete several reasoning and instruction-following tasks. Mercury 3 correctly fixes the registry’s donor-count and compliance calculations, answers a physical reasoning question, identifies a quote’s author, and generates multilingual brand names. It takes longer and struggles more with tool use on a complex request to create an interactive tattoo simulation, but produces a functional single-file page. The reviewer considers it a strong result for speed and instruction following, while noting limitations in creativity and visual artistry.

Key points

The registry demo has a bug: a donor with 40 births across five clinics is incorrectly shown as having six of 10 available family slots.
Mercury 3 fixes the issue, updates the compliance display, and identifies seven donors over the limit.
The video describes Mercury 3 as a diffusion-based language model that produces and refines tokens in parallel.
Inception claims the model runs at 1,068 tokens per second on GPUs and has a one-million-token context.
Mercury 3 correctly reasons that a marble remains on the table when a cup covering it is carried away.
It correctly attributes the quoted passage to Upton Sinclair.
For a multilingual naming task, it covers the requested languages and follows the specified language-name format, though many suggestions are literal translations.
The model takes considerable time and has some difficulty with tool use on the complex interactive tattoo-simulation task.
The resulting page is functional, but the dragon and thigh illustrations lack detail and realism.

Chapters

Start	Chapter	Summary
00:01	The donor registry bug	The presenter explains the registry’s family limit and shows a subtle bug in its donor counts and compliance indicators.
02:05	Introducing Mercury 3 and fixing the app	The presenter outlines Mercury 3’s diffusion-based approach and reported specifications, then shows it correcting the registry bug.
04:12	Reasoning, attribution, and multilingual tests	Mercury 3 answers a physical reasoning question, identifies Upton Sinclair as a quote’s author, and generates names across multiple languages.
07:22	Building and evaluating the tattoo simulation	The model takes longer to build an interactive tattoo page, which works but has noticeable limitations in visual detail and artistry.

Quotes

Mercury produces and refines tokens in parallel — 02:36

the answer looks correct, as you can see. — 03:40

That marble is on the table. — 04:43

a real solid model on instruction following — 09:25

Concepts and entities

Mercury 3, Inception, Hermes, Lineage, Fahad Mirza, modelos de linguagem baseados em difusão, raciocínio físico, Upton Sinclair, GPT-6 Luna, Claude Haiku 5.5

Full transcript (original language)

00:01 You are looking at Lineage, a national donor family registry. Fertility clinics in seven countries report every birth conceived from a donor, and the registry folds all those reports into one family count per donor. This is a real-world application. The policy limit is 10 families per donor, and uh donors at or over the limit must not be given to new recipient families. Why does this matter?

00:32 As I said, it’s a real-world application. The reason it matters is because every extra family means uh more half-siblings who may unknowingly meet or partner later. In December 2025, new reports described a donor carrying a cancer-causing gene variant whose sperm was used for almost 200 children across 14 countries. The limit in this app is a fictional demo policy at the moment, and data is synthetic. Now, look at the screen.

01:04 The dial here on the left says 100% compliance. Donors over the limit shows a green zero. This one towards the right. So, this is one of the bugs. Plus, if I just come down, you will see that this donor, which is let’s say D 2207, it has 40 births across five clinics, yet the ring says six of 10 available. So, something is wrong in the

01:35 application. Nothing is crashing, no error is appearing, but there is a subtle bug, which we are going to give to this model, which is Mercury 3. And I already have configured this model with my Hermes agent, where it is running through API at the moment. So, let me give it the goal. Where I’m asking it at high level to get this fixed and then give us a working application.

02:05 Now, while it works, let me introduce you to myself. I’m Fahad Mirza and I welcome you to the channel. I recently also have started this free weekly AI newsletter, which you can subscribe at fahadmirza.substack.com. You can see that we have been covering these inception models for quite some time. And they say it is their uh the latest one is their most performant model. So, this model is, as I said, from inception and it is a diffusion-based language model.

02:36 Most LLMs write one token at a time. Left-to-right diffusion models work differently. Mercury produces and refines tokens in parallel, a bit like how image diffusion refines a rough picture into a clear one, as you can see. That is the reason for its speed. Inception says Mercury 3 is um around 40 times larger than their first Mercury model and runs at 1,068 tokens per second on GPUs. It has a 1 million token

03:08 context. The price is also a bit low, but you can check it out on their model card. They claim it has beaten GPT-6 Luna and Claude Haiku 5.5 on their benchmark. Yes, Claude Haiku. Which is not that impressive even to say, but anyway. Um it is available through their API, which I’m accessing at the moment. So, let’s see how it does with this first test. Okay, so it says that it has finished the goal. And if you look at the thinking, that is very interesting. You

03:40 know, just like in a fashion of diffusion model, it has um it hasn’t really gone that deep, but still it managed to actually give us quite a good grounded and very pinpointed answer. So, there you go. It already refreshed. I didn’t even need to refresh. And the answer looks correct, as you can see. Now, the donors over family limit are seven, and this is a policy compliance. And if you look at now this D2207, it is now over limit. So, the model is quite good and fast.

04:12 Let’s do another test. And now let’s do a scientific reasoning test. If you look at this prompt, what we are testing is specifically whether the model tracks physical objects through a sequence of steps instead of pattern matching. So, the prompt is a marble sits on a table with a cup flipped over it. The cup is carried into a microwave, and the model must say where the marble is. So, you can also take your pick, but we will see if the model gets it right or

04:43 not. There you go. The model, you see how quick that was. Let me just take you up. And, you know, hardly any thinking at all, and then it has given us the answer, which is correct. That marble is on the table. And then it is going step by step. Now, look at these stages. So, how quick, how fast, and how accurate. Then it, you know, this is the main part where it is talking about inertia and all that scientific reasoning around suction,

05:14 adhesion, or the rim catching it from the below. Really good answer. Next up, let me test the factual recall and attribution. So, I’m giving it this prompt, and I’m asking it who this following quotation belong to. Now, let’s see if the model can name the right author of a famous quote without being fooled by a popular misattribution. This quote is from Upton Sinclair, which it has got it right. How good is that?

05:48 Next, I’m checking multilinguality at scale and also instruction following. Like, can the model cover every single language in a long test? And I’m asking it to come up with best possible brand name for a new eco-friendly water bottle in each and every one of the languages listed below. And here you go. This is the answer. Please check your language and let me know, too. As far as I can see, it has covered every language on the list. And it kept the exact language name format. That is the instruction

06:19 following we were testing. It also picked one clear idea, spring and pure water, and built the whole brand family around it. Which makes the set feel designed rather than a random. Um You know, you can argue that weak spot might be the creativity, like many entries are simply the word for spring. So, they read as translations more than brand names. But again, I think there are few standouts like Wellspring, Aqua Viva. And then also, I’ll just show you one

06:50 thing. Like this gibberish, Gluggle. So, it is quite cheeky. It also flagged the real languages like Gutnish, Balochi, and few others as best guesses that need a native speaker check, which is an honest touch. So, if you are a native speaker, please let us know in your comments. And finally, I’m asking the model to draw me a self-contained file showing a simulation of a tattoo. And there are various requirements around that. It should be a Japanese-style dragon tattoo

07:22 wrapped along it and few other things as you can see on your screen. And this is where model is taking now considerable time. It is thinking hard and long. Installing the tools, trying to use them. And this is what happens with diffusion models. As soon as the things become hard, complex, and long horizon, they start struggling a bit in terms of speed and tool calling and that is exactly what is happening here. It is sometime it just goes in circle then, you know,

07:53 gets back onto the track and then start using the tools. Anyway, um so far so good. So, let’s see how long does it take and what it produces and I will show you. And Mercury 3 has built something from a single prompt in one file. Let me copy it and open it in front of you and we will check it out together. Mhm, there you go. So, what do you think? It drew a skin tone thigh shape as you can see with soft shading a serpentine dragon with a head, claws,

08:25 and spine ridges and a working control bar, too. So, let me increase the set link. Is it doing anything? It is healed, healed, healed. Okay. There is a subtle change. You see? Which is good. Good works. And can I change in color? B&W, yep. You see? Black and white. Maybe it is flipped, but it is good. It is really good.

08:55 I think this is the first good one I have seen in a while from a model. Um but yes, there are a lot of flaws. Like, if you look at the dragon, it just like, you know, seems like a thin snake or a lizard sketch than a detailed Japanese dragon with no real scales, no flowing whiskers, and the thigh is a plain a blob uh rather than a realistic leg. But again, I think um it has followed the structure of the brief well. Uh artistry is not there, but again, this is not an image

09:25 generation model. So, I would say uh a real solid model on instruction following and getting a working interactive page in one go. So, functionality is also quite good. It is not a frontier model, but still I think for a diffusion model with this speed and this quality, it is not a bad version of Mercury from Inception, especially if you compare it to their previous models. Let me know your thoughts in the comments. If you want to help out the channel, please go to the home page and become a member or from the first pinned comment.

09:57 Thank you for all the support.