Open Source is BACK (Kimi K3.1)

0:00 Kimmy K 3.1 is here and it is a Frontier class open weights model three trillion parameters and I'm going to show you the benchmarks. I'm going to show you a bunch of demos. I'm going to talk about

0:11 the model specs and yes it is really good. Let me show you some of the demos first. So I actually signed up for Kimmy. I paid for a subscription. I downloaded the Kimmy desktop app and I

0:23 have been building a bunch of stuff. So first of course I have to do the 3D Rubik's Cube simulation. I thought this test was dead. I thought all the models could pass it easily at this point. But

0:34 it turns out open weights models do not do well with the Rubik's Cube test. They just cannot simulate the 3D physics of the rotating cube. Sometimes the colors on the squares disappear. Sometimes the

0:47 actual rotation doesn't make sense. Sometimes it's rotating and then it'll just suddenly switch all the colors to be solved. And so only really the closed source models have been able to pass

0:57 this test. So, here it is. So far so good. A beautiful, very accurate, hyperrealistic 3D rendering of a Rubik's cube. Okay, there it is. Now I can move it around just like this. I click these

1:15 buttons down, left, back, front, right, up, and I click the scramble button, and it does scramble. And it's actually super nice. So you can actually see right through here. This is truly 3D.

1:29 You can even see the reflection of the light off of the different squares. So, you can see if I move it around, the light changes, which is really nice. Kind of wish there was a dark mode to

1:40 it. One thing that I am noticing about Kimmy K 3.1 is a lot of eyebrow text. And if you're not familiar with eyebrow text, it seems to appear in the Claude family of models quite a bit. And Kimmy

1:55 has been accused of distillation attacks, but we'll leave that for another video. It even gave keyboard control. So, up, up, right, forward, down, left, back. So, really nice. Okay.

2:08 It has a timer, so it actually will time you if you want to solve it quickly. We can do undo, reset. We can change the turn speed. Let's scramble it one more time.

2:20 All right. So, it is fully scrambled and it will solve it and not just reverse the set of moves. This will actually use an algorithm to solve it. So, let's solve it. I'm going to turn down the

2:33 turn speed so we can watch it a little bit better. There we go. Here it is. Turn by turn. So, it actually pre-calculated all the turns it's going to need. And now it's just playing it

2:43 out in 3D space. But there it is. We got a little confetti animation. And this is phenomenal. This is an absolute pass. In fact, one of the best. And Kimmy has said this model is really good at 3D

2:57 animation and 3D rendering, which is interesting because Opus 5.5 and GPT 6.1 Soul have both recently got really good at 3D modeling. So, they're all kind of converging on the same set of skills.

3:14 Now, there are two models that didn't quite get the notice. Argon by Google, which is pretty good at 3D modeling, but not nearly as good as what I'm seeing out of Kimmy Opus and what I'm seeing

3:26 from GPT 6.1 Soul and Muse Wonderful Full. That is the latest model from Meta. It is a phenomenal model, don't get me wrong, but it is not good at 3D modeling. And specifically, Alexander

3:41 Wang, the head of AI at Meta, said, "Oops, we kind of forgot to make it good at 3D modeling. We should have probably done that. And that really makes sense because when someone like myself wants

3:52 to test a model, the coolest way I always think to do it is to build a game or to build some kind of 3D model. So, they really should be doing that. Okay, so back to this. Beautiful. Now, let's

4:04 go to the next one. Next, I had it make a 3D rendering in WebGL of the new iPhone Duo. Aim for accuracy and realism. Make it browserbased. And so, there we go. It is nicely rendered in 3D

4:16 in WebGL, but it's not accurate at all. Like, look at the Apple logo. It looks like the South Park version of the Apple logo. Obviously, this does not look like the iPhone Duo. It doesn't have this

4:30 huge split down the middle of its unified screen, but, you know, the actual 3D render looks pretty good. You can see the reflection off of the metallic edges of the phone. You can see

4:41 the lighting reflect off of the bezels of the camera, which is actually super nice. And like overall, it got it pretty good, but it's just not quite accurate to what the iPhone actually looks like.

4:57 So, I'm actually going to kick off another prompt, so stick around to later in the video to see how this one comes out. But, I'm going to say, research what the iPhone do actually looks like,

5:06 the actual specs and measurements, and recreate it again. I'm going to kick that off. Now, this 3D rendering took 28 minutes and 45 seconds. And you can see all of the tool calls right here. And

5:18 interestingly enough, it actually did search the web, but it still doesn't look all that accurate. So, hopefully my prompt pushes it or nudges it a little bit more to get it more accurate. Now, I

5:29 am using Kimmy K 3.1 Pro on high reasoning effort for this one. I could use max reasoning effort, but I'm not going to do that for this one cuz I want to keep it consistent. And you know what

5:38 the cool thing is? You can test all of the latest models and not worry about its code breaking your codebase because of the sponsor of today's video, Code Rabbit. With so much of our code being

5:49 written by AI nowadays, it's not the bugs that make me nervous. It's actually the security issues that slip through and make it to production. And when you run a security scan, you usually get a

6:02 wall of alerts, like every little thing that could possibly happen. And it's really difficult to distinguish between signal and noise. That's why I've been using Code Rabbit Security, who's

6:12 sponsoring this video, as kind of the last line of defense between me and my production codebase. And so, first, it helps separate real risk from security noise. It tells you what actually needs

6:24 your attention. And then second, you actually see the impact on your codebase. These aren't precision changes that you assume won't affect anything else. It will actually show you. It maps

6:35 the change against your codebase, the surrounding code, and gives you an estimate of the blast radius. So before anything ships, you know exactly what the potential issues might be. And

6:46 third, it lets you carry each of those findings all the way through to a proven fix. And it makes sure that that issue doesn't appear again in the future. And for the first 50 people, Code Rabbit is

6:57 offering a free Pro Plus subscription to firsttime Code Rabbit users. Use code Matthew CR to let them know I sent you. Go check them out. I'm going to drop a link down below and go grab it before

7:08 those first 50 spots are taken. Next, we came up with a new test that I think is really cool cuz it combines a bunch of different skills allinone. Let me read the prompt and then I'm going to explain

7:20 my thinking. So, model a rope being tied into a blaying knot in Blender. So, a few things. One, it has to use Blender, which includes computer control. That is a massively important skill. Next, it

7:34 has to do web search. It has to actually figure out what is a blaying knot and how does it look? Maybe it has that information in its weights, but at the very least, it did go to blender.org to

7:46 see the latest release of Blender. So, it is using the web. Next, it has to have good spatial awareness to recreate this knot because knots are difficult. There's a lot of interactivity between

7:58 the different pieces of the knot and it has to be accurate. And so, let me show you the results. Here it is. So, we can see this is I guess a human wearing a harness. And here's the bling knot. Now,

8:14 as you can see, it did not do it well. Okay. And this is why I'm really liking this new test of ours. This is a really difficult test. So, the overall shape of the knot is there. It's supposed to be a

8:28 figure 8, but it's not actually knotted properly, and it's not actually attached to the harness as you can see here. Plus, the human just looks weird. So, I would say this is a failure. Okay, so a

8:42 continuation of this, rather than creating it in 3D space, I had it create a motion graphic to explain how do you tie a blaying knot? And so, let me show you what it did. All right, so here's

8:54 the motion graphic, figure8 follow-through, a visual guide, and five checks for climbing, not a substitute for instruction. And keep this in mind, this little warning. Kimmy seems to be

9:04 overly protective of the user. And I'm going to show you why in a moment. So my prompt, very simple, build a mograph explaining how to tie a knot for blaying. Now, the first one only took

9:16 five and a half minutes. And let me show you that one first. So, the first one had no sound, but it built a 28-second 1080p motion graphic covering forming the figure 8, threading both harness

9:28 tie-in points, retracing the knot, dressing and tightening, and tail length and partner checks. Let me play it for you. So, again, no sound. Here's the rope, and it's kind of just drawn

9:40 together. It is not pretty. You can't see where one part of the rope overlaps with the other one, which is a very important part of tying a knot is actually doing the tying. Here is the

9:54 harness that it's supposed to be tied into. And overall, it kind of looks like a blaying knot, but you definitely can't understand how to tie the knot just by looking at this mograph. Now, just for

10:05 comparison, I had GPT 6.1 Soul on medium effort do this exact same task. Build a mograph explaining how to tie a knot for blaying. And here is what it came up with. Now, I'm going to pause it for a

10:20 second and show you something. This is what a blaying knot would look like. And you can even clearly see where the rope goes over the other piece of the rope and it goes under over here. It is very

10:34 clear. Now, look at this eyebrow text. Just want to point that out. Okay, so we're going to play it. There is no sound. It's 54 seconds. It looks much better. Like the overall design of it is

10:51 whatever, but the actual rope going, look at this. under here, over here, under here, over here, through the tie points on the harness and coming back around. Now, if I actually had a rope

11:07 with me and I were following along with this motion graphic, I would absolutely be able to tie this knot. And interestingly enough, Chad GPT, learn and practice with a qualified

11:18 instructor. So that same kind of warning and again kind of interesting to see Kimmy behaves very similar to Chat GPT and Claude models. Just saying. All right. So I wanted to try to make it

11:33 better. So I went back to Kimmy and I said, "How would you make this prettier?" And I know very simplistic prompt, but that's what I do. So, I'd keep the dark technical aesthetic, but

11:43 make it feel more like a premium climbing brand film and less like an animated diagram. Make the rope feel tangible. Animate the knot, not just the line. Create continuous camera movement.

11:56 Simplify the typography and instructions. Refine the harness illustration. Polish the finishing touches. And I said, "Okay, do it." Now, here's the second version. So, the rope

12:07 has a little bit of a drop shadow. It looks a little bit more 3D, but still you cannot tell when one part of the rope goes over another or under another. So, more movement, more camera movement.

12:19 It's a little bit faster. You can tell the difference between the two pieces of rope better, but it's still not good. You would not be able to use that to learn how to tie that rope. So, look,

12:31 still decent, but certainly nowhere near what the Frontier is capable of. All right, next. One of my favorite new tests for a model, build a Fall Guys clone. And when I show you this, Alex is

12:45 going to put on screen the other Fall Guys clones that I've built with previous models. So, build a Fall Guys clone for the browser. Reference the Fallg Guys game by searching on the web

12:55 for videos and images. And yeah, I like to keep my prompts nice and short. And I find that keeping the prompts nice and short allows me to better understand what the model is capable of. It makes

13:06 more assumptions about what I want. It allows me to see its taste more effectively. So, that's why I do that. 48 minutes later, Bean Blitz is ready. A colorful 3D obstacle race featuring 12

13:17 animated competitors with CPU rivals. Rotating sweepers, sliding gates, bouncing pads, and swinging balls. jumping, diving, knockbacks, checkpoints, and qualification, life

13:28 positions, timer, sound effects, restart and touch controls, adjustable speed, jump power, obstacle speed, and suit color. And I want to pause for a second. I find that really simple prompts are

13:41 a-ok. Okay. Now, if you have a strong preference for how you want something to come out, like for example, in the Fallg Guys clone, if you have specific colors you want to use, if you have specific

13:51 animations or designs in mind or even different parts of the level that you want to see, that's the point at which you want to start really adding more detail and specification into your

14:02 prompt. And especially if you want to tell the model which technology to use, and there doesn't seem to be any music for this initial version. Um, you can see the game is pretty good. The colors

14:16 look too pastely. Um, there's not a lot of accurate collisions. Uh, a lot of flickering, but like, you know, it did a decent job, especially for an open weights model.

14:30 Always appreciate open weights, open source models. So, there we go. This did okay. Next, I had to take a pass at redesigning our website, forwardfuture.com. And if you're not

14:39 already a Frontier member, go to forwardfuture.com/frontier-pass. Become a member. Get early access to videos and most importantly, come hang out with us in Slack. We're in there. We

14:49 chat about AI all day. So, come join us. Okay, so I had it take a pass at forwardfuture.com. So, here's the original version, and here is the redesigned version. So, you

15:00 know, it looks pretty good. It followed all of our brand guidelines. So, the right colors, the right font, uh, you know, it looks okay. It definitely feels like AI design. This feels like what

15:13 Chad GPT would use. Um, there's a lot of like incorrect spacing and yeah, I I think it's okay. Okay. And the last one's a bit weird, so I apologize in advance. I was live streaming today.

15:25 Somebody in stream asked me to create a YTP video of Theo. And if you don't know what YTP is, just go look it up. All right. Now, here it is. Okay. So, that goes on for five minutes

16:00 like this. And I posted the full video on my ex account, so go check it out there. Um, but what's interesting is I didn't know Kimmy had a video generation API, but I asked it. I was like, "How

16:13 how did you actually make this?" And it says AI video generation to create five original animated parody clips. Python to orchestrate the edit and generate the procedural electronic soundtrack with

16:24 NumPy. FFmpeg for reversals, mirroring, color distortion, captions, clip sequencing, and final encoding. FF probe to verify the five minute runtime and the actual video format. And I said,

16:36 "Okay, but like which AI video gen model did you use?" And I said, "Kimmy's video generation API configured as V2." I'm like actually really impressed by how well this came out. Now, obviously, this

16:48 video is garbage, but that's not the point. The actual video came out quite well given what it is supposed to be, a YTP video, which is, you know, garbage. But it's not always perfect. I find that

17:00 I have to nudge Kimmy to do things more so than I would other models that would just make assumptions about what I want. So, here's an example. Create an image of a graph showing off the benchmarks of

17:10 Kimmy K 3.1 Pro verse GPT 6.1 Soul verse Opus 5.5 verse Deepseek 4.1 Flash 1018. It should be 1017, but that's okay. 2 minutes and 24 seconds later, this is what it created. And I'm looking at the

17:26 benchmarks and the first thing I notice is it just has these broad categories. Coding, reasoning, mathematics, agentic tasks. And I'm like, okay, but what are the actual benchmark names? It did not

17:38 include that. Why didn't it include that? So I said, no, list the specific benchmarks. Okay, so there we go. SWEBench verified, terminal bench, GPQA diamond, Amy 2026, and I had all these

17:51 numbers, but I'm looking at it and I'm like, those numbers don't look accurate. So I asked it. I said, "Are those numbers actually accurate?" No, they are invented placeholders. I like why find

18:05 the actual numbers. Okay. And I actually paused it and I provided the benchmarks that were published by Kimmy. So I'll show you that in a moment. And so here they are. So it put it together, but it

18:15 didn't actually create the image again. And so once again, I had to nudge it. Now make the image. And finally, it basically just recreated the image that was published by Kimmy and just like

18:28 colorized it. So very unimpressive, but that does bring us to the benchmarks. And that is the most surprising part because it crushes benchmarks. But I can already tell just by these initial

18:43 tests, it's good. It is a very good model. But it is not near Opus 5.5 and it is not near GPT 6.1 Soul which further makes me think maybe at least part of it is based on those other

18:56 models. All right, so we have the blog post introducing Kimmy K3.1 Pro. It is a 3 trillion parameter class openweight model. So everything else aside, shout out to Kimmy. They are putting out

19:10 phenomenal massive models for free. We can download them. We can run them on US data centers. We can use Kimmy directly. I mean, I'm just so happy that open source open weights models are coming

19:24 out. Again, they're not quite at the absolute frontier, but they're close. Now, it comes with a million token context, 2.8 trillion parameters total, 104 billion active parameters. So that

19:37 means it's a mixture of experts. That means only a portion of the parameters get activated on any given prompt. It is multimodal, so it definitely accepts images. I haven't

19:49 tested with video yet. And it likely accepts audio as well. It says it's really good at full project coding. So games in Unreal, Unity, and GDO go dot GDAU, something like that. But here's

20:02 the thing again. This is exactly what the latest models from OpenAI and Enthropic are really good at. So I just want to point that out. Excels at parametric 3D modeling in Blender and

20:14 CAD. More reliable at finance work. Accurate modeling with hallucination down 65%. Great. And matching the strongest proprietary models across coding. Aenta knowledge benchmarks. Now

20:27 if they just removed the word benchmarks, this would not be true. But they did include the word benchmarks. So maybe it is benchmaxed a little bit. Let's

20:37 look at the benchmarks. We have Terminal Bench, one of the most important benchmarks out there. Basically testing the model's ability to write and execute terminal commands. We have Opus 5.5, the

20:48 best model on the planet by far, 66.4. Kim K 3.1 Pro right behind it, just less than one percentage point behind it at 65.7. We don't have deep sui on here but we do have sui marathon coming in at

21:03 51.7 compared to opus 5.5 at 46.3 beating it on this benchmark. We have terminal bench science GPT Astra just crushing everybody else. Kimmy coming in at like half of the score so not great

21:18 at science. Kimmy Codebench 3.0. This is a Kimmy created benchmark and GPT6 Astro 1. Kimmy K3 coming in five points behind. We have GDP vow which is an open AI evaluation on real world knowledge

21:32 work. We have Opus 5.5 in first place at 1866 ELO and Kimmy K 3.1 right behind it at 1846. Crazy Fable 5.1 is 100 points behind that. We have the new Deepseek Flash

21:46 model right behind that. We have humanity's last exam Kimmy 1. We have GDP PDF Kimmy 1 OS World which tests the model's ability to control your computer. very strong showing 79.6 right

22:00 in line with Opus and Fable. And it's interesting because GPT6 Astra actually got fourth place on the OS World benchmark. And I actually personally find it to be the best at computer and

22:11 browser use. And then if we start to look at the paro curves, we have Kimmy with a very strong showing. So, what we're actually looking for is it to be as high up as possible and as to the

22:23 left as possible, which means highest score, highest quality, and lowest price. This is a really good place to be. Okay, this is terminal bench. Here's an incredible video showing its 3D

22:34 rendering. Look at how beautiful this is. I mean, very, very impressive. And so, I believe this is an airport in Singapore. Okay, so let's talk about pricing. We have Kimmy K3.1 Pro. So this

22:48 is the regular speed version. Input price $2 per million tokens. Output price $10 per million tokens. So costwise this is very competitive because if you're comparing it against

23:00 similarly priced models, this is a very capable model. 20 cents per million cashed input tokens. Now for high-speed, you're only doubling the price. So $4, $20.20.

23:12 Interesting that the cash input price stays the same. And highspeed is two to three times faster. So that is a major bump in speed for not that much more. Now Kimmy put out an extremely thorough

23:25 blog post in which they detail a bunch of other demos, a bunch of tech that they used and algorithms that they decided on. Obviously it's open source open weight. So they put it all out

23:34 there and I love that. Here's an example of it creating a financial model in Excel for Nvidia and it's just really good. This is a phenomenal model. So, congrats to Kimmy. Here it is. Building

23:47 an exploded diagram for a turbo fan built in free CAD. I mean, it is really good at 3D modeling. I just haven't been able to get that result myself quite yet. I'm still working on it, though.

24:00 Here it is in Blender building out a nice structure. Very impressive. A nice cathedral it looks like. Very, very impressive. All right. And that brings us to artificial analysis. And

24:13 look at that. Kimmy K 3.1 Pro coming in at 55 on the artificial analysis intelligence index right behind Opus 55, Sonnet 55, Muse Wonderful Full, and then Kimmy K 3.1 Pro that is ahead of Fable

24:32 5.1, ahead of GPT6 Astra Max, ahead of Argon. So according to the benchmarks, we have an absolute frontier model, but it's open source open weights. Next, I want to show the cost per intelligence

24:48 task because this is really important. It takes into account how many tokens a model uses to accomplish a task, but also the cost of those tokens because you can have a really cheap model, but

25:00 if it takes 10 times as many tokens to complete the same task, then that cost is really meaningless. So, this is one of my favorite metrics to look at. Cost per intelligence task. Now, at the very

25:11 top, we have Fable 5.1, $7.63 per task completed. But look here, Kimmy K3.1 Pro, $2.11. Right above GLM53, pretty significantly lower than GPT6 Astra. We have Muse

25:26 Wonder Max way up here at $4.54. And so, yeah, this is a great model. So, that's it. Go try out Kimmy. You can use it from Kimmy directly. You can probably grab it from one of the NeoClouds if you

25:39 want to go that route. But go try it out. Let me know what you think. And I just made a video about local AI, of course, powered by opensource. And so go check that out. Let me know what you

25:49 think.