Full transcript
0:00In this video, we're going to be diving
0:01into another new model and getting into
0:04my test to see how good is it for
0:06creative writing. This is one that a
0:08number of you uh mentioned in the
0:10comments of my other videos that you
0:11wanted to see me test. And I was
0:14actually very pleasantly surprised when
0:16I started digging into all all of the
0:19different prompt results. So the model
0:21that we're covering today is Kimmy K2
0:23thinking which is a newer model that
0:26comes from a it's a Chinese company and
0:30it's an advanced open reasoning model
0:32that is actually pretty good. So let's
0:35dive in. So this is the model here in
0:37open router giving us some information
0:39about what it's all about and uh there
0:41are a lot of interesting things here.
0:43First of all it let's take a look at
0:45price. Now, the price that it has up
0:48here is a little cheaper than most of
0:50the prices that I actually see it down
0:52here in uh in open router. So,
0:55theoretically, it should have 40 cents
0:57per million tokens of input and $1.75
1:02for million tokens of output. But here,
1:04realistically, I see it at 47 cents and
1:07$2 respectively. So, not that much of a
1:11difference, but that's just something to
1:13note here. And depending on where you're
1:15getting it from in Open Router, because
1:17Open Router pulls from various different
1:19servers, you might even get a little bit
1:21more expensive, like $2.50.
1:24But that said, this is still incredibly
1:27cheap. It is not as cheap as another
1:30model that I recently reviewed, Mr.
1:32Small Creative. Uh Mr. Small Creative is
1:3510 cents per input and 30 cents per
1:37output, which is insane. Uh, but this is
1:40kind of round about the the same level
1:43as Gemini 3 Flash, which is the smaller
1:46Gemini 3 model. Uh, that one goes for 50
1:49cents for input and $3 for output. Uh,
1:52but that all of this pales in comparison
1:55to uh Gemini 3 Pro, which is $2 per
1:59input and $12 per output. And and Gemini
2:033 Pro isn't even that expensive of a
2:05model as models go. Uh the claw models,
2:09most of them are more expensive than
2:10that, especially Opus. Uh and yeah, I'd
2:15say like the Gemini 3 Pro prices here
2:19are pretty standard for like a main hero
2:25model, if we're going to call it that.
2:28Um also the context in Kimmy K2 thinking
2:32is pretty high. Now up here it'll say
2:35262,000
2:37tokens of context. Down here it's
2:39different. Um this one that was a little
2:42that was 47 and2 has 131. Um here it's
2:46262. So again it depends on like what
2:49server you're getting it from. But all
2:51in all that's a pretty decent context.
2:53Like I could run pretty much all of my
2:54automations with that amount of context.
2:57U so I'm I'd be happy with that. Uh so
3:00very very cheap affordable model.
3:03uh from Moonshot. And let's actually get
3:06into some of the results for the prompts
3:09that I give it and see how well it does.
3:11I'm going to be uh naturally probably
3:14comparing this a little bit to Mistral
3:15Small Small Creative, which I just
3:17recently reviewed as well, so I can't
3:19really separate the two in my head. Uh
3:22but hopefully you we'll get a good sense
3:25of how good this model is. So, let's
3:27start with the log line prompt where I
3:30ask it to give me 20 epic epic fantasy
3:32log lines in Blake Snider's style. Let's
3:35read through a couple of these. A
3:36mathematically gifted farm boy discovers
3:38he's the chosen one destined to defeat
3:39the dark lord, but when he calculates
3:41the prophecy is 99% wrong. Both sides
3:44hunt him to make the math work because
3:47his failure is the only thing that
3:48actually saves the world. Interesting.
3:50I'm not sure that's actually
3:52a a feasible story there. When a mind
3:55healer is kidnapped by the dark lord to
3:57cure his nightmares, she discovers his
3:59evil is a magical parasite, but curing
4:01him will shatter the fragile alliance
4:02that only his tyranny holds together,
4:04forcing her to choose between his sanity
4:06and civilization survival.
4:09Now, I think there's actually a decent
4:10story in there. Uh maybe not exactly
4:13like it's listed here, but I could I
4:16could work with that. A financially
4:17inept dragon hires a human accountant to
4:20manage his horde. All right, so we're
4:22getting into like fantasy comedy here.
4:23But when she realizes the horde is the
4:25kingdom's treasury, she must choose
4:27between embezzling for the crown or
4:29being eaten only to discover the dragon
4:31is the rightful king to pose for being
4:32too generous. Now, that one's a little
4:33scattered and all over the place, but
4:35overall, just looking through these, I
4:38would say these are slightly better than
4:40what I was getting out of Mistress Small
4:41Creative. Maybe not quite as good as
4:43what I was getting out of Gemini 3, uh,
4:46but certainly okay. And if you gave it a
4:49little bit more u to go on in terms of
4:53the right tropes you want to hit for
4:55your specific subg genre and all of that
4:56like here it's definitely going towards
4:59the the more generic tropes like chosen
5:02one dragons and you know dark lords and
5:06stuff like that. Uh if you were to get a
5:08little bit more specific with your
5:09guidance you might be able to get better
5:11stuff out of it. Um but I would say it's
5:15solid. Not great, but solid. All right,
5:18moving on to the outline test prompt
5:19where I give it a big long outline
5:21template and a story premise and ask it
5:24to write a whole a whole outline from
5:26it. And this one it does okay. Now,
5:29because this actually is a reasoning
5:31model, it does actually set up a number
5:34of events that take place later. My
5:36biggest problem with the Mistro small
5:38model that I reviewed earlier was that
5:41it really felt like each chapter was
5:43kind of siloed and didn't pay much
5:45attention to what happens in the other
5:47chapters and so it felt very choppy and
5:49stilted. Uh this one is not quite so bad
5:52and this one actually finished the
5:55entire thing. I actually went through
5:57and and got through the entire outline.
6:01Um, I will say it started by giving lots
6:04of information here at the very
6:05beginning.
6:07Um, and then by the time we got to the
6:10bottom, you know, these chapters were
6:12getting a little bit smaller as you can
6:15tell. Um, and that's typical of a lot of
6:18models. Even Gemini 3 does that. The
6:20cloud models do not do that, which is
6:22interesting. Um, but overall a pretty
6:25solid outline. Um, it's kind of hard to
6:28to really compare this one. All right,
6:30now we get to the basic pros prompt.
6:32This is really what interests me the
6:34most is how well does it actually write.
6:37So, let's go through. It actually wrote
6:39a fair amount for this particular
6:43scene. Um, so usually I kind of jump to
6:47this moment in the scene to to read it
6:49and see just how how good is it.
6:51Scattered figures dotted the ridge he
6:53had crusted an hour past. Their
6:54movements wrong in ways that made the
6:56eye re the eye rebel. two dozen at
6:59least, though counting them felt like
7:01counting grains of sand in a storm. They
7:03moved with that characteristic jerky
7:05motion, limbs snapped into place as if
7:07pulled by invisible strings, then
7:08freezing, then snapping forward again.
7:10Their skin, what little showed beneath
7:12ragged remnants of clothing, was the
7:14color of the wasteland itself. Gray as
7:16ash and dust, making them nearly
7:17indistinguishable from the rocks they
7:19navigated with such unnatural precision.
7:21They were close. Too close.
7:24Valdon's gate gates rose before him now
7:27carved from the mountain's own bones 100
7:29feet of iron wood and granite that had
7:31never been breached's heart hammered
7:32against his re ribs as he measured the
7:35distance maybe 300 yds to the gate house
7:37another 50 to the port cullis the math
7:39was simple and brutal even running as he
7:42was even with the last reserves of
7:43strength that desperation had unlocked
7:45the hungry would reach him first he knew
7:47the city's policy had helped enforce it
7:49himself in younger days standing watch
7:51while some of the other soul begged for
7:53sanctuary that would not come. When the
7:55hungry were within bowshot, the gates
7:57did not open. Not for anyone, not for
7:59any reason, but he had to try. And
8:02usually most of the models would stop
8:04there. It did go on a little bit
8:06further. Um, but not in a way where it
8:09was like going beyond the story, uh, the
8:13beats that I gave it. It definitely
8:16stopped at a good pace there. And it
8:19added in some details that I thought was
8:20really good. um like he had helped
8:23enforce it himself in younger days,
8:26standing watch while some other soul
8:27begged for sanctuary. Um you know,
8:30that's that's a little added detail that
8:31kind of fleshes it out, makes it more
8:33interesting. Uh but wasn't in the
8:36instructions about how, you know, we
8:38don't know anything about his past or
8:40anything like that. Um so I liked that.
8:43And overall, I'd say the pros is pretty
8:45decent. not a whole lot of AISMs and
8:48definitely pretty good description and
8:51all of that. Uh, I thought this was
8:52pretty solid honestly. Right up there. I
8:55would say almost this is almost as good
8:57if not just as good, maybe slightly
8:59better than Mr. Small Creative, which I
9:01reviewed recently. Um, now let's look at
9:04the complex pros prompt. It's the same
9:06scene, but uh I added a little more
9:09context like some examples of my writing
9:12to to mimic and stuff like that. uh more
9:15information about the characters and
9:16whatnot. So, let's see if it did any
9:18better. It did write a shorter amount,
9:20which is typical when you give it more
9:22context. Um, let's see what it wrote.
9:27Sound cut through the wind, not the
9:28moaning gust, but something else.
9:29Something wrong. Haldrid looked back.
9:31Figures dotted the ridge he had just
9:33crossed. Two dozen, maybe more. They
9:34moved with that sickening jerk stop
9:36motion that turned his stomach. The
9:38hungry, their gray skin, nearly
9:39indistinguishable from the rocky terrain
9:41they navigated. Too close. Far, far, too
9:43close. Valdon's policy was carved into
9:45every citizen's memory. No entry when
9:47the hungry were within 500 meters.
9:49Quarantine protocol. Safety of the many
9:51over the few. He turned toward the city
9:53gates. A desperate calculation burning
9:55through his exhaustion. The policy
9:57didn't matter. He had to try. I I almost
9:58think I like the first one a little bit
10:00better. Probably because it spent longer
10:02to do it. This does feel like it was
10:04following my style a little bit better.
10:06I tend to have a lot more of these like
10:08smaller sentences in there. So, I um I
10:11do think it did okay.
10:14at that part of it. Still okay. I think
10:17the the first one was better. But let's
10:19move to the dialogue pros template. This
10:21is a prompt where I ask it to write a
10:23different scene with a with a dialogue
10:26aspect to it. So, let's go in there and
10:29read a little bit about uh from this
10:30scene. Another man died at our gates
10:32tonight, Lean said, his voice cutting
10:33through the formal silence. He stopped
10:35before the throne, not bowing. Never
10:36bowing. Uh, I like that little detail
10:39right there. Never bowing. The second in
10:41a month, Father. I watched the hungry
10:42take him while your guards held me back.
10:44Store's expression didn't change. The
10:46law is the law, son. The law. Lean
10:48laughed, a bitter broken sound. You are
10:50the law. You sat on that throne and
10:51wrote those words. You could unwrite
10:53them with a breath. Your highness Argal
10:56interjected his tone smooth as oiled
10:58silk. The laws exist for a reason.
11:00Valdin has stood for 300 years because
11:02we do not compromise on survival. To
11:04open the gates when the hungry are upon
11:06us would be to invite them into our very
11:07streets. Surely even you can see the
11:09wisdom in that. So, this is not bad. I
11:12do think that when I compare it to Mr.
11:15Small Creative, I think Mr. Small
11:16actually did better on the dialogue than
11:18this one, interestingly enough. Uh,
11:20there are a few moments in there that I
11:22really do like, like his the notion of
11:25not bowing, never bowing. That's a good
11:28like deep point of view moment there,
11:30uh, that you don't get out of a lot of
11:31AI models, stuff like that. So, I do
11:34like that. Uh, and it's certainly not
11:36bad. I just think that in comparing it
11:39to Mr. Smell Creative, the the latter
11:42might have done a little bit better.
11:45All right. Now, for my editing pros
11:49prompt,
11:51uh, for this, I give it a the same scene
11:54written by a much more inferior AI
11:57model, uh, in this case, Claude 2, and
12:00ask it to improve it just to see if it
12:02can actually, you know, improve on the
12:04writing or is it better to just rewrite
12:06the thing? Uh, and this is what we got
12:09here. Father, his voice cut through the
12:11incense thick air. We we will speak of
12:13the gate. King Store sat on the obsidian
12:16throne. His silhouette a fracture of
12:17darkness against the rose windows dying
12:19light. He did not lean forward. He did
12:21not frown. He simply waited as mountains
12:23wait for storms to spend themselves. "A
12:25man died last night," Lean said, halting
12:27at the deis's edge. "Not quickly, not
12:29cleanly. He died begging for the mercy
12:30we denied him while you watched from the
12:32wall." "The law held," Store said. The
12:34words fell like stones into a deep well.
12:37"The law is yours to break." Lean
12:38Knuckles were white at his size. One
12:40word, one command, and the gates would
12:42have opened. He would have lived. All
12:43right. So, I don't think this is great.
12:45Um, it's it's definitely still an
12:47improvement on the original text. Uh,
12:50but I thought Mr. Small Creative
12:51actually did better at this and
12:53certainly Gemini 3 did better at this.
12:55Um, so some improvement, yes. Um, but
13:01certainly not the most that I would
13:04expect, especially when compared to some
13:06of the other flagship models from other
13:07com from other uh AI companies. All
13:11right, let's move on to ad headlines to
13:12see how well it did at this.
13:15Uh, I ask it for 20 Facebook ad
13:17headlines. Betrayed and cast from his
13:19mountain home. The truth he discovered
13:20outside was worse than death. They told
13:23him out the outside world was death.
13:24They forgot to mention the real danger.
13:27Zombies, hive minds, shadow demons, and
13:29one lie that doomed a kingdom. What if
13:31the monsters outside aren't the ones you
13:32should fear? All right. Um, I would say
13:34none of these are particularly great.
13:36Uh, Mist Creative was not great either.
13:40Uh, this doesn't really feel any better
13:42to me than that one. Uh, a fantasy where
13:45the world building is the mystery and
13:47the truth is the ultimate revenge. That
13:49one's maybe okay. Uh, but it's still
13:53kind of vague. Uh, not really a good
13:54hook. Room for improvement there. I
13:56definitely think Gemini 3 did pretty
13:58good at this one in comparison. All
14:00right, email newsletter prompt. So, this
14:02is where I give it the description of
14:04the book and ask it to write a launch
14:06email for it. After months of writing,
14:08I'm thrilled to announce that my epic
14:09fantasy novel, The Crown of Hunger,
14:10releases on December 21st. Betraying
14:12Cast from His Mountain Home, yada yada
14:14describes the whole thing. Pre-order
14:17now. Mark your calendars for December
14:1921st. Best my name. And that's all it
14:22is. And I got to hand it to it. I do
14:25like shorter emails. I don't know if
14:29that's necessarily something it was
14:31thinking about um that shorter is
14:33better, but I do think that sometimes
14:36shorter is better, especially in emails
14:38like this. I'd rather send out more
14:40frequent but short emails. So, actually,
14:42I think this is pretty good. But that
14:44said, because it was so short, there
14:46wasn't really much much of a chance to
14:48really show what it it's capable of. But
14:52on the whole, yeah, this is fine. I I
14:53don't really have any any issues with
14:55it. All right, last but not least, we
14:56have the SEO article super prompt where
14:59I give it a bunch of articles uh about
15:02how to write a fantasy novel plus an
15:03outline for an article and ask it to
15:06write the entire article. And I asked
15:08for 4,000 words. It gave me 2,300. So, a
15:12little bit missing the mark there, but
15:14more than what I got out of Mr. Small
15:17Creative. On the whole, just looking
15:19through this, it looks like it does an
15:21okay job, but it's certainly not
15:22fleshing things out as much as I would
15:24want it to. It's not quite the like more
15:29just like a bulleted outline of what the
15:31article should be that I got out of Mist
15:33Small Creative. Um, but it's not
15:36particularly
15:38great either. So, on the whole, better,
15:41but still not really great compared to
15:44others. Uh, right now the cloud models
15:46are my favorite for SEO articles. They
15:49do a really good job, even better than
15:50Gemini 3, which you would expect the
15:52Gemini models to do really well at SEO
15:55because it's Google after all. Um, but
15:58no, I find the CL cloud models to be
16:00quite good for that. Uh, and this
16:02definitely doesn't compare to that. So,
16:04on the whole, my thoughts on Kimmy K2
16:07thinking, it's it's a solid model,
16:10especially for the price. Uh, even
16:12though it's not quite at the level of
16:13Mr. Small Creative. It's still
16:15significantly cheaper than most of the
16:18flagship models. Does it compare in
16:21quality to those flagship models? I
16:23don't think so. But as a mini model, it
16:28it's probably one of the better ones
16:30that you could use for something that
16:32requires, you know, that that you want
16:34to be very very cheap. I I need to do a
16:37proper comparison of this versus Gemini
16:393 Flash, which is priced very similarly,
16:42but my gut instinct would suggest that
16:44it might even do better than Gemini 3
16:46Flash uh for certain tasks. So, on the
16:51whole, really solid model, definitely
16:53one worth doing. We are definitely
16:55moving into an age where it's not going
16:58to matter so much which model you use
17:01because all of them or most of them will
17:04actually be decent at doing different
17:06things. And it's going to come down a
17:07little bit more to your prompts and your
17:10your context. And uh in the software
17:14world, it's going to come down to how
17:15you can use these models in creative and
17:17inventive ways to really save people
17:20time in a way that actually works. Um,
17:23so I I think this is exciting. I think
17:26we're moving into a really exciting age,
17:29uh, where, you know, we don't have to
17:31fight over which model is the best
17:32because they're all going to be good
17:34enough, right? And it's going to be more
17:36about you and your what you bring to the
17:39table as a human that is going to set
17:42the books apart. So, those are just my
17:45thoughts here on Kimmy K2. Thinking is
17:47solid model, uh, but not, you know, not
17:50quite at the level of a flagship model.
17:52Uh, and I hope that was useful for you.
17:54I'll see you in the next video.