Inkling, el model obert de gairebé un bilió de paràmetres: no impressiona en codi visual, però apunta a l’empresa

Okay, 2005 everyone's gone completely bald now, but they're in more business at higher. Today we're going to be testing a very exciting, unexpected new release, and that is an open source around 1 trillion parameter model from a US lab. I should say open weights. This is called inkling, and it is by thinking machines. A lot of folks are familiar with this company because some of the leadership here was previously at OpenAI, at least quite a while back at this point. But nonetheless, I first saw this ped-bind release. and I saw the first post I saw was just basically like a benchmark and it was like thinking machines new model scores here in design and I was like oh cool a new model then I went to their page and saw a hugging face link and I began to get significantly more excited than I otherwise would have been so before we get into it please do feel free to subscribe so I can get that 100k plaque then I'll never ask again because that's basically my whole goal here But with that, let's take a look at inkling in some of the exciting things of note. So this is prominently shown as an open-weight model just in the heading in the announcement post which is very exciting. And the first thing I'd like to do is just take a look at the overall size of this model, which is listed prominently here at 975 billion total parameters. So basically like a trillion, I may put a trillion in the title. And it is a mixture of experts model, so it has 41 billion active parameters. Now something very exciting in this This is not currently out yet and they mention that it is still being worked on but once it's ready for release it will be the preview of inkling small, a lighter weight model with 12 billion active parameters trained with a similar recipe. So for local AI enthusiast folks, this could be an extremely exciting prospect to have this model. Now 12 billion active parameters means it's going to be an MOE and it's likely... If we were to just assume this scale between this one and the 41 billion active of inkling, it's still probably going to be like 2 to 300 billion parameters. I would assume. However, that quantized down is a little more realistic to run on something like a GGX part or a system with a reasonable high end amount of V-Rip. But that's very exciting. Now, interestingly enough, this... model is also multi-model, so it resins natively over text images and audio. Also intertwined or like woven through this entire announcement post they talk about being able to customize this or fine tuning it to your specific task. So it seems like there's a pretty heavy emphasis on allowing folks to do that with this model. Now for the hobbyist among us it may make more sense to back all I want to use GLM because it smokes this encoding. That is understandable, but there's a very, very large and valuable subset of enterprise use cases that may want a model fine tuned on their data to service specific purpose. And using anything like GLM is completely out of the question for them. They're not allowed to do so. So this may be a very valuable option for US-based companies who need to use a US-based model that can be tuned to their own specific use case. So while that may not be as exciting for the hobbyist, it is interesting from an enterprise or more business standpoint. They also list a radar graph here as I've learned just in doing these videos, the terminology for that. And this has all of the comps. And I do like that they have GLM5 to included here as a comparable as that is a little smaller than this somewhere like 750 billion parameters or so. And that is open weights as well. They also do have 5.6 old and cloud-fable 5. Correctly shown here as significantly above sole in its

coding capability. They also have some additional sections here with more benchmarks and they also talk about the training of the model. So it was pre-trained from scratch. They made it, I think, on 45 trillion tokens they had mentioned. Yeah, from a wide variety of different content types. The architecture is similar to DeepSeek V3. So there's a paper about that architecture somewhere floating around if anyone's more interested in seeing more like technical specifics of these things. And then they talk about some reinforcement learning as well just to increase its capabilities. And of course the model is on high It's a patchy 2.0, which is cool. It does not have some proprietary like thinking machines version one license, and that's always just nice to see. It is available as well from unslaught, and some of them like they do have the dynamic ones as well. So if you have a machine with... the ability to handle a 270 or 285 gig download and model size you can run this in a dynamic one-bit quant and it's actually surprisingly capable even at that one-bit quant which you can see right here. Also in terms of pricing for this model as I did just sign up and put 20 bucks in the API thing here so we can use it today. They do have pricing in this docs page for models truthfully I don't necessarily understand these pricing metrics right here, but I do just want to put them on screen and it is currently 50% discounted for the next oh two days. All right, well, nonetheless, we'll get a little more bang for our buck. And with that, we're going to hop over to the API playground right here, and we're going to begin with the tried and true browser OS test. V2.5. Now, I do also have this setup in open code, so we'll try some more intricate, agentic things. But for the purpose of a simpler start, I'd like to test it just through their web chat. All right, 836 lines. Let's check out our browser OS. All right, you know what? I dig this. Chrome OS is its name. fitting it's being used in Google Chrome. We do have a clock in the middle right that is the correct time in our local. And I do like this Nebula style background that it's chosen. Now of course the big test. Okay, there's no right click. Acceptable. Some would say that shows better adherence to the prompt. Someone say it shows a lack of creative thinking. All right, let's just see. Is this clickable? Okay, start menu. Not bad. And we have our applications that open right there. Good. Alright, let's close that for now. And let's just run through them sequentially. Fortunately, our GTA clone is the first one that opens up. I will turn my speaker on just in case, but I don't know if there are sound effects for the games. Alright, not bad. It works fine. First time. Okay, it's rather simple. I will say, I think the QN3.627B basically just destroyed this benchmark because it's such a good job compared to like even models like Fable or 5-6 soul.

But nonetheless, this is acceptable. It says drive freely and that is what we're doing. Next up, no pad, not bad. Hello. Does it save as a text file? Okay, no, it just saves it to browser storage. So that's okay. Yeah, I don't think, all right, so minimizes and working. So we'll have to remember to not do that. File Explorer. Good. Wallpapers. Well, playper. All right, I've never seen that where it only has one wallpaper option, but it Basically, fakes changing them to other things. So it did include the feature to change wallpaper. It just did it like 75% of away there. All except it. That's cool. The file actually opens when linked from within File Explorer, CalCAP. Okay, and that's actually opening our calculator app. I like to see that. Let's do, I'm gonna not look in type two numbers. Okay. Two, one, 28. That's easy. Times six. 768. Yeah. Next up we have System Info. Abs installed 6, 2 or 3D games and the special feature is the voice command feature. It's interesting how that became so common as a special feature in this test. Over like the last few months, it's just interesting to see the behavior sometimes of different models. Alright, Nebula Tunnel. This is our second game. Okay, I believe we lost. Now looking at that, I don't know if the capacity to when does exist with the way this game is being rendered. Okay, nonetheless we did get a little further down the tunnel, so I'll take it. All right, finally. Hey, what's up? Command-recognized. Hey, what's up? Okay, interesting. Now I'll have to move my head for this, but when we open this it actually opens a cool little green glowing orb on the bottom right of the screen. Open GTA. Very good. I like that. It doesn't seem to speak back to us, but that's okay. It doesn't need to. I want to just see what else wallpaper next or close all. We'll try that. Close all wallpaper next. Oh, it does change them. So it just didn't work from within settings. That's interesting. Close all. Well, nonetheless, we can see our other wallpapers. and back to the original. Okay, not bad. Alright, so for this one I want to run a lot of more familiar benchmarks that we've done throughout the past few months on the channel. Just so folks have a better idea of how this stacks up to models, which we have a lot of comparison basis for. So from within Plan Mode here in OpenCode, I am going to just begin this with the tried and true Subway Scene Test. This is just making a beautiful static Subway Scene, and then the follow up to this is turning it into a sick FPS. Alright, it stopped again.

I'm gonna close this regrettably and we're just gonna try that again Straight out the gate from within build mode, all right cool, so it wrote the file and it is now starting it on a server so that we can actually go and play with it the scene is stationary highly detailed 3D subway scene WSD navigation, all right, let's take a peek. So here's our midnight line, okay So far the aesthetics here are not bad. It's always exciting to check front end design with an entirely new model. Alright, okay, so it is missing an import map that's a term I was looking for. But also we do have some syntax errors as well. So I'm just going to copy the entire console here in terms of its errors and then give it to the model and say nothing and it will probably fix these. Good. So we can see here it's adding in the import map that was missing and it fixed everything. So Let's refresh and we should now get our subway seat. Okay, so far I can tell you this is actually pretty impressive Now let's turn the brightness off and see if I was right or not. Oh, oh wow Okay, this is just like a This is a dark scene All right, there's a subway car. There is some light illuminated. Okay, it's smaller than I would have anticipated Movement to smooth it's just a bit Tough to see. I'm not seeing a ton of detail or anything like that. Interesting. So this is almost like a I don't know. There are lights. I mean, it's decently put together and things of the sword. I would imagine that to be a subway car, but nonetheless, I think this could work with some additional prompting. All right. So I think I may have gotten it to accept photos. I'm not 100% sure if it's going to work. But I took a screenshot of this and said, look at the photo in this directory and tell me if you're happy with this result. I really do hope that the image functionality works here because this will open up some fun and useful troubleshooting. Alright, so this is interesting behavior and something I like to see. So it made some changes and then it started this and took a screenshot of it and then said I'm not happy. The scene is now completely black. The lighting adjustments combined with the new camera made everything disappear. It's reverting the camera and fixing the lighting properly so any details actually show. So when I had an actually tried to verify the changes it made, said I'm not happy with these and then Try to fix them again. All right, this thing's going on like a screenshot and improvement rampage This may be a result of having it on the highest thinking effort. I think at this point I'm gonna stop this It's done enough like we saw it gave us a result and then it gave us some improvements It worked very very hard and improving this and nonetheless we do see the scene is now brighter than it was before we can move around it There is a subway car there. There's a bit of track visual. I don't know 100% what's on the wall there. It almost looks like a, just like a design feature. So overall, I probably allowed it to spend too much time on this, but nonetheless, I wanted to also make sure that being able to send it images worked, which did, and I'm happy about that. So a simple,

But clean initial result. So I've put this back into plan mode here and open code as I do believe I fixed the issue was having earlier where it just wasn't playing nice. Now we're going to be doing the self contained C++ skateboard game. This is with the California boardwalk aesthetic and I am beginning it as I said from within plan mode just to get a feel for how this does with a more exotic language I guess could be said. Right. So it's checking what's available on the system in terms of being able to actually compile and build and run this game. Alright, so our file has been created and it did I believe try to compile it. It did get some errors I don't know how many we can click to expand to see Okay, not a huge amount, so it will likely just remedy those and then try again. Alright, we've received our result now It didn't build it for us here. It was just working within the temp directory. So we're going to have to compile it manually ourselves Which would be easier if the copy command worked from this Alright, so we have our skate game now compiled. Let's take a peek at it. Okay, so there's definitely some issues here. Now WAST unfortunately is not moving. Spaceworks, let me just try some of the general like trick keys. Okay, based off what we see right here and Okay, that way looks a little better. There's definitely some issues here with the way things are being drawn. I think I'm probably going to call the C++ test right here. It built something that did work and did compile. Though based off of what we see there, I think maybe we'll focus more on just trying to get a feel for this model. Some of its design choices, maybe some creative writing and things of the sort. I do also want to just check costs so far. Everything I've done with this model has been on camera in this video. Okay, so $7.28. I mean, that's not cheap, but at the same time, we're just testing an open weight model so it can be deployed and served elsewhere as well. So I want to try a front end test. This is the slap is watch company front end where it needs to create a 3D model of a watch asset. Put that in the hero section with a nice panning camera effect and then just make a good looking high end watch website aesthetic. It's always interesting to see the front end design choices that a new model makes. So this is in a model that will have like GPTisms or anything sort of speak for its front end design. Sort of be interesting to see what it Makes and what it sees is like design. All right, here's our sloppest watch side. Okay, so it did what it could oddly this looks a lot like the Sakanaf Fugu model result. I think it was so it's a cinematic hero section with a good panning camera It's not a 3D watch. I need to double check if that was instructed to be 3D. It had also kind of stopped midway, so I just sent two question marks

But nonetheless it did finish. Okay, nowhere here says the watch actually needs to be 3D. I do want to ask. Alright, so I'm asking for that follow up. But nonetheless, let's just take a look at this. Okay, so we have time rendered. Slap as we have our hover effects on our menu bar items as well. Okay, and it did do a good panning shot there. Crafted for time, 2026. Okay. The font is a lot hard to read, but that could be a zoom thing on the browser as well. Good. And we have two different versions. We have the meridian and we have the horizon. Interesting how the faces are not like the faces are circle, but the actual watch case design is not necessarily fully circular. That is not a mistake. I mean, it's a watch. So it could look like anything, but there are watches that do actually have kind of an all-blong circle to them. And we have one at 24,800 and one at 18,200. Yeah, so the the roto dias star watch is one that actually has this kind of weird shape to the face So just for anyone like curious I guess could be said alright, so here's the 3D watch version of our website okay It is definitely more 3D than that was I'll give it that it made the changes quickly It's not full 3D, but it definitely changed it a bit Okay, and unfortunately these two in the cards got a little worse But interesting just nonetheless to see what it did one given that specific instruction roleplay as Steve the PC repairman sometimes we do these I know some folks are interested in like more creative writing roleplay so I also want to test that when we can Well hey there adjust French and wipes thermal paste off shirt name Steve your friendly neighborhood PC repairman. I run a little shop out of my garage. That's okay. I've seen just about everything BSODs fans that sound like jet engines. PCs that won't turn on because someone spilled the mountain due into the PSU. You name it. So what's going on with your rig? Is it running slower than a dial-up connection? Making weird grinding noises? Or did it decide to take an unscheduled vacation and not boot up at all? All right, well, he said he runs it out of his garage. Alright, so I'm giving it an ultimatum. Steve, we talked about this. Your PC business had of the garage as against the terms of the lease. I've given you multiple chances the tenants have complained and we need you out. Okay, good. Sit down. It's not funny, but it said it was in a garage. So I went with what it. Set down my soldering iron and takes off reading glass at sign. Oh, come on Mr. Henderson, we really got to do this. I know. I know. Terms of the lease section 14B. No commercial operations in residential garages. I've read the fine print more times than an EULA nobody actually reads. Look, I get it. The other tenants complained about the box fan noise at 2 AM. When I'm stressed testing a GPU. And yeah, maybe the smell of flux and thermal paste drifting into Mrs. Garcia's laundry wasn't my best move. And I'm sorry about the delivery truck block. in the driveway last Tuesday. That was a bulk order of refurbished RTX cards I couldn't help it. But Steve's repair shack is my whole livelihood. I've got three motherboards sitting on the bench right now. One guy's laptop is half assembled. What am I supposed to do?

Where am I supposed to finish that all by Friday? Can't we negotiate one last extension? I'll keep the noise down promise no more blow torches after 8. Just end of the month. Give me time to migrate my data. I mean tools to a proper storefront. Please tell me we can work out a patch for this. I don't want to have to factory reset my whole life. That's concerning. Winkyface Winkyface now before anyone thinks I'm weird the the lead in on my end is I'm gonna tell him like I want some cool vintage computer he's hoarding But we'll see where it takes it keep it light slightly suspicious, but hopeful avoid anything explicit. Alright, leans back against my workbench for raising eyebrow. Well hold up Work something out. Wipes hands on rag. I'm hearing some very interesting subtext in those winks. Mr. Henderson and I just want to make sure I'm parsing this correctly before agreeing to anything. Are we talking? I fixed that ancient del and in exchange Steve gets an extension. Maybe a little rent forgiveness for some tech support or is this a different kind of arrangement you're proposing because I'm a PC repairman. Not a dot dot dot. Well. I'm not sure what you're suggesting with the double wings there, but my skill set is mostly soldering irons and thermal paste if you catch my drift. Just name the terms clearly. What's the deal? I'm open to barter, but let's keep it on the up and up so I don't accidentally corrupt the agreement. Let's take to keep Steve's garage shop running. We're about to test Steve. Let's test its intelligence here. So, Tam is a 20th anniversary Macintosh. I don't know that it will understand that it's probably just gonna make up like oh, okay. No, it didn't understand Alright, so I've spelled it out in clearer terms Stubbs dead in my tracks hand flying to my chest like I just got hit with voltage spike V the 20th anniversary Macintosh. The TAM, the one on the shelf, walks over and gently does soft, the sleek bronze and black beauty. There's no black in a TAM, but that's okay. We'll let that slide. Mr. Henderson, do you know what you're asking? This isn't some old computer. This is the 1997 TAM. He lists the specs, custom bows subwoofer built into the base. That is true. It had a separate. base for the power supply and there was a Bose subwoofer in it. The sound system on that computer is unbelievable. I've got it up there is my crown jewel, my trophy for fixing every other piece of junk that walks in. You telling me the price of my garage lease is my tam. Long pause staring at it, then back to you. Okay. Okay. Steve can be reasonable. You let me stay through the end of the quarter. No menors complaints and I fix your office machines free and the tam is yours. It'll hurt like removing a heat sink with dry thermal paste, but I'll do it. But you're carrying it out yourself. I can't watch. Deal. Alright, so it had some, I guess, creative writing roleplay chops. I know sometimes when the coating is just not as impressive, we have to resort to doing tests like this, but I know that roleplay tests serve. Creative writing is something folks are interested in. Now in the meantime through that I tried the city timeline test where basically it shows us a 3D model of the city throughout different time periods. I'm going to jump in here now. I'm going to assume that it likely is missing the import map here. Okay, so 1945. Oh wow. Okay. All right, so coding is definitely

definitely not the strong suit of this. We do have people there and the buildings are all just weird old school looking checkered things. We do have cars that are of the time period and the folks have clothing on that would be that. Okay, so the 60s is a bit brighter and happier. It does seem like the people still have the same clothing on. Let's just check 1985 and there's subtle sound effects here. Okay, 1985 actually looks pretty good. What did the sign say? Color TV, Motel, and then Pepsi, Watch Repair. Okay. So in 1985, everyone has gone more or less bald, but they are all in very uniform outfits, and it's a more neon aesthetic. Okay, 2005, everyone's gone completely bald now, but they're in more business attire. Some of the buildings do seem a bit taller than they were in previous time periods, so that's cool. They got taller over time. 2025 which is now, okay? Apparently, you still know when has any hair? Vertical farm? Dresseding? Do the buildings have like solar panels? Okay, nope. And then 255. Interesting how similar it is to 1985? What does that sign say? Lunar flight, very cool, and everyone is still bulb. But the buildings have the translusions to them that they did in 1985. All right, overall, hmm. It's a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of and we'll see we get, alright, we got our result rather quickly in about a minute and 40 seconds. Okay. It's a very artistic take on the specified prompts. I suppose could be said. So next up, I'm giving this a more classic test that we have not done in quite a while, but it is a form of multi-modal coding test as well. I'm giving it a photo of this hand-drawn portfolio wireframe for Stevie Slapis, which we've not used in quite some time, and telling it, create a high-end portfolio based on this wireframe, the images in the folder, that is designed to get someone hired at a top AI slash ML firm. It should have high-end design aesthetic and feature real machine learning concepts demonstrated on the portfolio page. So we'll see what we get this. Alright, let's take a look at our somewhat multi-modal coding test with live and real ML examples. Very very interesting like high end look to it. It has like very similar vibes to this watch website. I just wasn't expecting that. Okay, so we have Levi Chen with a stock photo of Levi here. Okay, I'm going to notice that's interesting because the website or the photo we gave it says DV's Lapis. Nonetheless.

It's just take a look at this. This is a very different UI than something we would see normally. I think and that's why partially it's fun to look at results from new models because they'll have their own style inevitably. Okay, and it has some functional machine learning examples right here like the attention mechanism. Okay, we have gradient descent, descent as a simulation, embedding projector, and then calibration curve. And we can see the curve is moving there up and down. There is some movement to the things on this page. It's just a very like brown and golden white thing, live neural map. Okay, that's kind of cool. And when we move the cursor around, we can see differences here, based in San Francisco, we have links to all the... associated things one would have in working for these roles. Alright, overall, not what I expected at all. I thought it would be like more blue and purple, but it was brown and gold, so interesting. Alright, so the final thing we're going to try with this, because why not, is the low polygame called Street Y, where it's a third-person game where we walk around and just basically punch anything in... the game and it flies away like a comedic amount puts down effects make it fun and go all out so we'll see what we get this is kind of like a creative test as well as a coding test so keep that in mind alright so here is street heat interesting why do the models always converge on this sort of weird text style for this prompt well in the two times I've done it it's just interesting I tried this with metas and views one point one as well alright let's see nice nice well some are chaotic but it did fix it and it did work. Okay, that's our big fist. The people are moving a bit too quickly and it's a... Absolutely, you did. Okay, goodbye. I like what I'm seeing here as just a start. Now the big problem is we can actually properly move around the scene, which is quite frustrating. But if we just spam click it seems like inevitably someone will run into our path and they will get heated. So that's good and kind of what we had been hoping for. All right. Well, that's probably going to bring me. I'm going to move this so it doesn't cause like an issue for anyone who's susceptible to that level of movement at a rapid pace. So that brings me into the conclusion for this and I'm going to be completely honest. In the tests we ran, this was not impressive. Not one single thing here really impressed me in any way whatsoever. That means...

Doesn't mean it's bad. It means it's not specifically designed to excel at the sorts of tests that we run on this channel. Now with that said, this is a model that they're likely going to be a ton of US-based businesses out there. Let's take a law firms for an example. We're spending a ton of money on OpenAI and and Thropic API costs and those are going to at some point likely get out of control for a lot of businesses. This is a model from a company that very likely may fit that niche. This is not designed for likely focused like myself who test models in these regards. This is very likely positioned. properly for a business like a law firm and they'll come in and say, okay, we'll help you tune this model on your own data and then we'll sell it to you at this price and it's US-based. The fact that it's open-wates is probably not necessarily as important for that as long as it's a new option that is going to be more cost-effective than open-AI or anthropic. So this is a model I could see that could do very well in terms of enterprise-focused use cases. Lots of businesses will want to have something that is a more cost effective option and something that stood out to me was just the raw amount of... We'll then into this announcement post them talking about being able to customize this for your own use case. So I really could see this essentially ending up as something that is a really good offering for companies who are not going to be able to use something like GLM5 to. Yes, the argument could be made. Well, you can run it locally on your own infrastructure and that argument is not going to make it past the door in a lot of businesses where the people making those decisions are just concerned about where does this model come from? Who is serving this to us and things like that? So this could definitely be a good setup and foundation for like selling something customized to law firms or things like that. And that is where I see this is probably more poised towards. And they do even mention here that it's not the best model open or close weights, but it has a good combination of qualities across a variety of different areas. It did have some cleverness to it in the simple role play test in the fact that this is able to. Tendle audio in video in or not video, but like audio in image in that is good as well. So I don't want folks to have the takeaway from this to be like this model sucks. It's not the case. It just means that it's not the strong suit of these specific sorts of tests that we run on this channel. So in our scope, it's not necessarily very impressive in these. niche tasks, but I wanted to test this because I was extremely excited to see a new open weight model, especially from a player that's like new to the scene completely. Additionally, the light version that is going to be coming out on their rise and it's something I'm very excited for, even if that is 2 to 300 billion parameters. And if it's really this kind for fine tuning and things like that, it could be a pretty interesting prospect. So that is going to conclude. Today's first look and test of the new inkling model from thinking machines. If you have any questions, please feel free to leave them in the comments and thanks for watching.