Vast-10M Hands-On: Does 10 Million Tokens of Context Hold Up?

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: Vast-10M Hands-On: Does 10 Million Tokens of Context Hold Up? · Channel: Fahd Mirza · 11:00 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Prompt and context engineering, Software engineering with AI, AI models

Summary

The video tests Waltropy’s Vast-10M model, which uses a modified attention mechanism to offer a claimed 10-million-token context window. Across four hands-on tasks, the model produces a playable-game attempt with a missing character, a decent but imperfect HTML tattoo illustration, a strong answer based on Australia’s Corporations Act, and a multilingual naming table with repeated ideas. The reviewer is positive about its performance on the legal corpus and generally finds the responses useful, while noting weaknesses in reasoning, visual detail, and consistency. He also says more technical information and open-sourcing or open weights would be welcome.

Key points

Vast-10M is presented in Flash, Pro, and Medium versions, each with a claimed 10 million tokens of native context.
The model’s VSA mechanism replaces the attention mechanisms in three existing models rather than starting from scratch.
The video’s first test asks the model to build a 3D “floor is lava” game; after almost an hour, the game’s character is missing, so the reviewer considers the test unsuccessful.
For a detailed biker tattoo, the model creates a working single-file HTML page with several requested elements, but the illustration falls short of the requested intricacy.
On seven volumes of Australia’s Corporations Act, the response is well structured and the reviewer spot-checks a cited section as accurate.
The legal-task reasoning loops and mixes up file names, though the final answer flags some information it could not find instead of guessing.
The model produces a table of names in 81 languages and flags uncertainty about Gutnish and Tinya.
The claim that none of the names share a root or idea is false; the reviewer spots repeated concepts, including “love” and “longing.”
The reviewer says more information about VSA and making the model open source or open weight would be welcome.

Chapters

Start	Chapter	Summary
00:01	Introducing Vast-10M and the first test	The reviewer introduces the model, its claimed context length and VSA mechanism, then starts a multi-tool task to build a 3D game.
03:07	Testing the generated game and tattoo	The game runs but has no visible character, while a later HTML tattoo illustration passes as a first attempt but lacks some requested detail.
07:16	A legal question across a large corpus	The model answers a question based on seven volumes of legislation with a useful, checked response, despite looping and file-name confusion in its reasoning.
08:51	Multilingual creativity and final verdict	The reviewer assesses a table of dating-app names in 81 languages, points out repeated ideas, and sums up the model’s strengths and limitations.

Quotes

Very big claim. Let’s see if it survives contact with my terminal. — 00:01

The problem is that there is no character which is jumping. — 04:41

The response is actually quite good. — 08:18

Because you see at least, you know, I can spot easily nine rows just mean love. — 09:55

Concepts and entities

Vast-10M, VSA, DeepSeek V4 Flash, DeepSeek V4 Pro, GLM-5.2, Hermes, Fahd Mirza, Australia’s Corporations Act

Full transcript (original language)

00:01 10 million tokens of context. And if you look at this chart from their technical report, this model beats Anthropic’s Fable 5.1 at 1 million tokens. Very big claim. Let’s see if it survives contact with my terminal. The model we are talking about is this vast 10 million from Waltropy. What they have done is they have taken the attention mechanism in three open models: DeepSeek V4 Flash, DeepSeek V4

00:31 Pro, and GLM 5.2, and replaced it with their own called as VSA. I’m going to talk more about it later. But for now you need to remember is the result is this model. And it comes in three flavors: Flash, Pro, and Medium, each with 10 million tokens of native context. And we are going to test it out thoroughly. This is Fahad Mirza, and I welcome you to the channel. Please also subscribe to this free weekly AI newsletter at fahadmirza.substack.com. Now, a full disclosure. Waltropy gave me

01:04 some API credits. This is not a sponsored video, though. I did ask just like I ask you guys to become a member. And as usual, they also said that they wanted an objective review. So, no dice. So, it seems that I’m just going to die poor. But lucky for them, objective is the only kind we do at this channel. So, if this model stumbles, you will hear about it. If it flies, you will hear about that too. So, let me take you straight

01:34 to my terminal, where I already have configured this model with Hermes agent. And first up, this is what I’m going to test out. Now, this is quite a challenging prompt, and I think the only model which gave me decent response for this was Fable. So, what this prompt is doing, it is asking the model to build a complete 3D floor is lava game from nothing. Blender scripts for the furniture and lava go. 4

02:05 for the game play with a timer, start screen and game over screen. And you can see that it has started building. So let’s see if this vast 10 million can plan and run a long multi-tool agentic job on its own and finish with a playable game without any rescue from me. While it works through it, let’s talk about that VSA. Now I really couldn’t find much information around that VSA, but after going through their technical

02:36 report, it seems that they did not build a new model from scratch as I mentioned earlier. They took three existing models and swapped their attention mechanism with their own, which is called as world trophy scalable attention or VSA. But their claim is that it stretches context to 10 million tokens and also scores higher at shorter context with no extensive retraining. So, really good stuff and there are a few paragraphs here and there, but I think it would be really good if they could get us more information and it would be

03:07 swell if it they could really open source this or open weight this model. The model has given me the result as you can see and from the output so far it looks strong. Um headless result with some tests too. So, let’s check it out. I’m going to press control C. You see it is it has taken almost 1 hour and let me quickly show you what exactly files the GLB one it has created. So a lot of stuff it has created as you can

03:38 see. But anyway, instead of doing that, let me run that game in front of you. I’m just doing it live so that um you see exactly what is happening. So now I’m going to press space to start. So, but where exactly is the person where I’m jumping? That is the main thing. I couldn’t see anything. Now, there are some assets which are missing, so maybe I will just do this control, see? And maybe I will

04:11 first import them in the headless. This is a path. So, it is importing those. And then let me rerun it. So, import is done, so maybe that is why it was missing that. And now I’m going to rerun this stuff. So, let’s paste it again. And let’s wait for it to get launched. Let me press the space bar here. There you go. So, you see, I’m jumping. The jump

04:41 works. I will just press enter again to restart. And I’m moving here. It also moves, but as soon as I move, it doesn’t really show anything. the problem is that there is no character which is jumping. So, that is the main issue there. I don’t think so this has passed this first test, but other than that, it was able to create those characters and everything, but if it has taken 1 hour, I think it should have done at least a complete job. Anyway, let’s do

05:11 another test. For the second test, I asked the model to design an intricate biker tattoo on a realistic forearm as one self-contained HTML file to see how it handles detailed visual work from a single prompt. This was my prompt, as you can see, quite detailed, quite rich, and I have explained everything. And this is where after thinking long and hard, it has created this biker tattoo file, which I have just opened in front of you in the

05:42 browser. Now, if you look at here, to be very fair, it delivered a working page with no reported errors and several requested elements are there. A skull, crossed pistons, three roses, and a red flame cluster, a mandala at the top, and a ride or die banner on a dark uh background. So, this is not an image generation software. This is just a plain HTML file. The skin tone and shading give a believable hint of depth, and the roses

06:12 are the most polished part. But, I think where it falls short is world-class, museum-quality. If you check out my prompt, go back to that screen and pause and check. The skull is simple and lacks the visible uh winged crown and wings. This snake looks like a loose black coil with no clear scales, uh really. Um so, that is the weakest part, in my opinion. The banner text is also overlapping a bit, but that is fine. I think that is um borderline nitpicking. Overall, I think

06:44 um it has listed quite a few things, but it has also struggled with few. So, a decent first pass illustration, but I would say uh bit short of intricate design requested, but I will, you know, mark it as a pass. And now, let me take you to their hosted platform. You can see that I have um uploaded seven volumes of Australia’s Corporations Act, about 4 MB of text, and not one of their curated demos, and

07:16 I’m asking one hard legal question that forces the model to connect sections across the whole act. Then, I will check every section number it cites against the real law. So, you see 10 million context window, so should be a real fun. It is sort of a needle in the haystack, too. So, let’s see what model comes back with. The model has come back with the response and it is quite an interesting one.

07:48 Um, if you you know, look at the reasoning and I will actually show you because I’m more interested in reasoning and how exactly it reached to that conclusion and response. It is actually looping a lot. It is checking and then rechecking which files are loaded many times and then it has also mixed up the file names like volume 02 versus volume 03. It has admitted that it is reconstructing a passage from memory

08:18 then says, yes, present in the text with no proof. But if I just scroll down and show you the response, the response is actually quite good. It is well structured. It has followed my groupings in the prompt and it has also flagged what it could not find instead of guessing or [clears throat] hallucinating, which is always welcome. And I have also sport checked sport checked the section 1841 on AustLII and you know, it has matched and I think the

08:51 answer is quite accurate, too. So, very nice. I’m quite happy with the answers accuracy. I think reasoning could be improved but other than that, it has done very well with this large corpus of data. And finally, let’s have a look at this prompt. Here, I’m testing its creativity, cultural depth and consistency across the whole list in one reply. So, I have asked it for a different culture specific dating app name in 81 languages with strict no

09:23 repeat rules to test creativity and as I said, language breadth and whether it admits what it does not know. And this is the response. Please also check out, let me know in the comments what do you think about your own language. The format is perfect. One clean table, all 81 rows, no extra text, and it honestly flagged two languages it was unsure of, Gutnish and Tinya, which is fine. It is quite rare. And then I’m just going to scroll through all the languages and in you know, it is done okay for

09:55 various of them. It has even given us some of the translation, too. So, pretty good. It’s closing claim that no two names share the same root or idea is false, totally. Because you see at least, you know, I can spot easily nine rows just mean love. Like in Pashto, Latin, Esperanto, and few others. Latin amor and Esperanto amor share the same root and longing repeats many times. But there are various strong picks, too. But still some look doubtful to me, so I

10:28 will take your you know, advice, especially around the low-resource languages. Um, let’s say the regional languages in India, maybe, or southeastern languages. So, let me know in the comments. But look, I think it is not a bad response at all from this model. So, that’s it. That is a vast 10 million model for you. Let me know your thoughts about the model. And again, if you want to help out the channel, please become a member by just clicking on this last link or in the first pinned comment. And please

11:00 subscribe to this newsletter. Thank you for all the support.