Full transcript
0:00In this video, we are going to dive into
0:02a new model that is specifically trained
0:05for creative writing endeavors. This is
0:08something that to my knowledge has never
0:10been done before in a public setting.
0:11Pseudorite does have their own
0:13proprietary model specifically for
0:15fiction writers, but it's kind of
0:17expensive and um if there were to be
0:20another option, I would definitely go
0:22for it. So let's go check out what this
0:24model can do and is it actually good
0:26enough to be used as your main workhorse
0:29when it comes to writing creative
0:31fiction. So the model in question here
0:33is Mistral Small Creative
0:37and uh it's first of all a very very
0:40tiny model as you can see from the info
0:43here. Um, this has a total context of
0:46only 32,000. And uh, unfortunately,
0:49that's one of the biggest downsides
0:52right out of the gate for this model is
0:54that it has a very low context window of
0:5732,000. Uh, I usually like to have at
0:59least 100,000. That's, you know, I'm
1:02going to be using at least that much.
1:04And when I was running this process
1:06through some of my uh when I was running
1:08this model through some of my
1:11automations, some of my automations were
1:13actually exceeding the context window of
1:1432,000. So this is not very big on the
1:17context. However, you can make it work
1:20if you're using a tool like Novelcfter
1:22to be very specific in the context that
1:24you're giving it. So you're only giving
1:25it just the right context for the
1:28specific scene you're writing. And
1:29hopefully that can help. But if you're
1:31giving it like your entire outline and
1:33your entire story bible and maybe even
1:35like the entire book up to that point,
1:38that might be overkill and it won't work
1:40with this model. Um, it's similarly its
1:43max output is the same about 32,000. But
1:46here here's the kicker here in the input
1:49price is 10 cents per million tokens.
1:53The output price is 30 cents per million
1:56tokens of output. That is insane. Okay.
2:01Uh that is a very very small model and
2:05very affordable. So if this can do even
2:09like on par with the heavy hitters like
2:11Gemini 3 and Claude Opus, uh those are
2:15in my opinion the two best models at
2:17this point in time for creative writing.
2:20But you know Gemini 3 be which is even
2:22the cheaper of the two. If we go look at
2:25those models, that one has an input
2:27price of $2 per million tokens of input
2:30and $12 per million tokens of output,
2:32which is significantly less than the 10
2:34cents and the 30 cents here. So, if we
2:38can get even close to the same
2:39performance, like it it'll be worth it
2:42to be using this model. And the nice
2:44thing about this is it's described here
2:46as an experimental small model designed
2:49for creative writing, narrative
2:50generation, role-play and
2:52characterdriven dialogue, uh general
2:55purpose instruction following, and
2:56conversational agents. So, this is
2:57really designed for any kind of creative
2:59endeavor. So, creative writing, uh if
3:02you do any kind of like Dungeons and
3:03Dragons type online roleplay with, uh
3:07AI, this could be a good model to use
3:09for that. Anything that uses dialogue or
3:12is conversational here. So, if you're
3:14doing uh like some kind of uh agent that
3:19has a conversation with the person, this
3:21might be a good one to look at,
3:22especially given how cheap it is. Now,
3:24as of right now, there is no MRO medium
3:26creative or MR Large Creative. Uh I am
3:29hoping to see one of those in the near
3:31future. This is an experimental small
3:33model, so maybe we'll see something more
3:35powerful in the future. But, let's
3:37actually get into my tests and see what
3:40I got out of it. and um and I'll let you
3:42be the judge of how good the pros is.
3:45So, first of all, this is I have a
3:48number of prompts that I go through with
3:49every model just to prompt them on the
3:52exact same basis. And the first one is a
3:54log line prompt where I ask it to give
3:56me a bunch of Blake Snideresque
3:59log lines for a epic fantasy series. And
4:03why do I do epic fantasy? Because a it's
4:06a genre I know, so I can usually
4:07understand if it's doing well or not.
4:10And B, I find the AI often struggles a
4:13little bit with the more archaic aspect
4:17of fantasy. Um, sometimes there there's
4:20an archaic aspect of it. Um, often when
4:23a model does well in contemporary
4:25genres, anything that's set in the past,
4:28it struggles with sometimes. So, if it
4:29can do the the past uh or fantasy,
4:32anything like that, it can usually do
4:34really well in most other genres. So
4:37anyway, uh let's go through some of
4:39these options that it gave us. When a
4:41disgraced scholar exiled for heresy
4:43after predicting the return of the
4:44eclipse god discovers the ancient
4:46prophecy was her own forgotten memory,
4:48she must recover the shattered sunstone
4:50relic before the church of the veil who
4:52burned her village for blasphemy uses it
4:54to permanent permanently extinguish the
4:56sun, only to learn the god she worships
4:59is the very darkness she's been
5:00fighting. Okay. A little a little bit
5:03scattered in its ideas, but there's
5:06there's something in there, maybe. Uh, a
5:08half orc mercenary raised to believe
5:10he's the bastard son of a warlord
5:12uncovers that he's actually the last
5:14heir of the Bone Kingdom, a dynasty that
5:16ruled through necromancy. And the Lich
5:18Queen, his aunt, has spent decades
5:19hunting him, not to kill him, but to
5:21crown him because the throne's magic
5:22only awakens for a willing sacrifice.
5:25Uh, that sounds a little bit more
5:27interesting. Still not nothing really
5:29fantastic there. So there's definitely
5:32some interesting ideas in some of these,
5:35but I definitely wouldn't say that this
5:37is anything particularly amazing. And
5:40that makes sense, right? Because this is
5:41not a thinking model. It is not a model
5:44with who who is particularly good at
5:46reasoning or anything like that. And
5:48this is more of a reasoning type task.
5:51It does speak very creatively and I
5:54think it does have idea like if I could
5:56somehow combine this with something that
5:58thought a little bit deeper. Um I might
6:02like it but so far I'm not being too
6:05amazed by anything that it gives me. Now
6:07the outline test prompt uh where I
6:10basically I give it a premise and ask it
6:12to outline a whole book based on that
6:14premise and an outline. I give it a big
6:16long outline. So this is one of the most
6:17intensive of the prompts that I give it.
6:19It's got quite a bit of context to it.
6:22And um I'll just tell you right now, it
6:25did not do a great job at this. First of
6:27all, it did not finish the book. It only
6:28got about halfway through the outline.
6:31And second of all, if you actually go
6:33through and read each chapter, uh what
6:36it wrote for each chapter, it's almost
6:38like each chapter is a very different
6:40story than the previous one. And it's
6:42not doing a good job of really
6:44connecting a through line through all of
6:46these chapters. It's like each one is
6:48sort of focused on that one by itself,
6:51then moved on to the next one and
6:52focused on that one by itself, and
6:54there's only a minimum amount of
6:57connection between the the different
6:59chapters. Um, so I wasn't that impressed
7:02with this, but again, this makes sense
7:04because the kind of models that you need
7:06to be able to produce a good outline
7:08require a lot of reasoning. It requires
7:11things like being able to understand
7:13foreshadowing and making sure that
7:15earlier chapters are for foreshadowing
7:17later chapters. And you generally want a
7:19higher reasoning type model to be able
7:22to do that well. I find Gemini does a
7:24good job. I find the Claude models, the
7:26most recent ones, do probably the best
7:28job at this one in my opinion. Um, but
7:32this model, it doesn't seem to be what
7:34it's designed for. It's not designed for
7:36thinking deeply about these things. It's
7:38designed more for the actual writing and
7:41the dialogue and stuff like that. So,
7:44let's go and take a look at some of
7:46those things. Next, we've have first our
7:49basic pros prompt, uh, which is just the
7:53it's describing a scene. It's like the
7:55beginning of a prologue of a of a book.
7:59And, um, let's actually get through and
8:01see what it says here. So,
8:04um, it describes the scene a little bit,
8:07the area, which they all do that. Um,
8:10and then we get a sound cut through the
8:13wind, a wet clicking noise like bones
8:15knocking together. Haldrid froze, his
8:17pulse hammered in his ears as he turned,
8:19his boots grinding into the gravel on
8:21the horizon where the wasteland blurred
8:23into the dark, gathering dark, figures
8:25moved. Not like men, not like anything
8:26living. Their limbs jerked in unnatural
8:28spasms, their gray skin blending with
8:30the ash and earth. the hungry. Dozens of
8:32them. Too many. His stomach twisted. He
8:34knew the rules of Vald in. The gates
8:36would not open if the hungry were this
8:37close, but he had to try. Um, so this
8:40definitely not bad. Like there's there's
8:42not much to this. I would say maybe we
8:44would want a little bit more fleshed out
8:46with this, but the pros isn't bad.
8:48There's definitely a reduction in AISM
8:52from what I've seen for other models,
8:54especially really cheap models like this
8:56one. Um, it's definitely, you know, it's
9:00it's solid. I would not say it's
9:02amazing, but it's solid. Okay. Uh, maybe
9:05average would be a good one, especially
9:07for, you know, when compared to the
9:09higher, beefier models, right? Then I
9:11have a complex pros prompt, which gives
9:13it the exact same piece of text, but I
9:16add much more to the prompt, uh, like
9:18style information. I give it an example
9:20of my own writing to emulate, stuff like
9:22that, just to see does it do any better
9:24when given a whole lot of this context.
9:27Uh, so let's go to the same section. Um,
9:30he didn't dare look back. Not yet. But
9:31the silence behind him was worse than
9:33the sound of pursuit. Haldrid slowed,
9:35his pulse hammering in his throat. He
9:36turned. The horizon wasn't empty.
9:38Figures moved there. Dozens of them,
9:40their jerky, unnatural gate
9:41unmistakable. Grayskinned, tattered
9:43rags. The way their heads twitched as if
9:44listening to something just beyond
9:45hearing. The hungry too close. Too many.
9:48His stomach twisted. Valdan's gates
9:50wouldn't open for him now. Not with the
9:52hungry this near. The city's policy was
9:53clear. No one in When the Dead Are This
9:56Close, but he had to try. I'd say that's
9:59a slight improvement on the previous
10:00one. Uh definitely a little bit more
10:03variety in the sentences and paragraph
10:06lengths and all of that. Uh but not a
10:09huge improvement, but still like decent.
10:11Like I I think this is pretty good. It's
10:14also for both of these. I should note
10:15that it's following the um it's
10:19following the the beats pretty well.
10:22Sometimes, especially with older models
10:24or some of the less powerful models that
10:26I've tested, it doesn't do a good job of
10:28that. It'll it'll like keep going after
10:30the end of the beats and kind of create
10:33create what it thinks happens next.
10:36And it didn't do that with either of
10:37these. So, that's a good point in its
10:39favor. Uh, then this is a different
10:42scene with some dialogue in it, so we
10:44can kind of test out what the dialogue
10:46looks like. The heavy oak doors of the
10:47throne room groaned as Lean shoved them
10:49open, his boots striking the stone floor
10:51with enough force to echo through the
10:52vaulted chamber. The scent of polished
10:55iron and old parchment hung thick in the
10:57air, mingling with the faint metallic
10:59tang of his own sweat. His chest heaved
11:01not from exertion, but from the fury
11:03coiling tight in his gut. All right, so
11:05this is uh it's okay. It feels a little
11:09bit AIish
11:12uh just because of how it is describing
11:14things, but I think this is still
11:16reasonably okay. Uh it's kind of hard to
11:19say with some of these things. But let's
11:21skip to the dialogue here. You storm in
11:23here like a man possessed, Argle said,
11:24his voice smooth as oil. What has you so
11:27agitated, Lean? Lean's hands clenched
11:29into fits. The gates again. Store
11:32finally lifted his gaze, his dark eyes
11:34unreadable. Explain the man at the
11:36gates. Lean bid out. The one who begged
11:37for entry. The one that hungry tore
11:39apart while we watched. Augle sighed as
11:41if disappointed by a child's tantrum.
11:43The law is clear. No one enters after
11:45dusk. The risk is too great. The law is
11:47the law. Lean mimicked. His voice
11:50dripping with scorn. He So things like
11:53that. His voice dripping with scorn. I
11:55feel like AI does little clauses like
11:58that a little bit too frequently. He
12:00stepped forward, ignoring the way the
12:01guards along the wall tensed. But the
12:03king makes the law. So tell me, father,
12:05why do we let people die? Store's jaw
12:08tightened because the alternative is
12:10worse. Is it? Lean's voice rose sharp as
12:13a blade. Because from where I stand, it
12:15looks like we're becoming monsters
12:16ourselves. Argos thin eyes, eyebrows
12:19lifted. Monsters? No, Lean. We are
12:21survivors. The hungry do not reason.
12:23They do not negotiate. They consume. If
12:24we open the gates for one desperate
12:26soul, we risk them all. So, definitely
12:29an improvement here on the dialogue. I
12:31think this is very good. Um, is it
12:33better than Gemini 3 or Cloud 4.5 Opus?
12:38I think that's debatable. I would
12:39[clears throat] have to really work with
12:40it for like multiple chapters worth
12:43going through and editing it to see if I
12:45liked it better. It certainly doesn't
12:48have any particular style that is sort
12:50of like pushing at me to be, you know,
12:54there's something about some of these
12:55models that you just you read enough of
12:57it and you're like it gets old after a
12:59while. And I'm not seeing anything like
13:01that. There's no unique style. It It
13:04feels a little bit more like Open Window
13:06Pros where there's where it just sort of
13:08lets you get into the story and hear it.
13:11Uh there are a few AISMs here and there,
13:14especially in the way it describes
13:15things, but overall really solid um for
13:18a dialogue scene. Uh, then I have it uh
13:21I have an editing pros template where I
13:24give it pros that was it's actually the
13:26same scene written by I believe it was
13:28Claude 2 way back in the day. Uh, and
13:31it's not great pros and I ask it to
13:34improve on that writing. So, let's read
13:38a little bit of that and see how it
13:40does. Father Lean's voice cut through
13:42the silence sharp as a drawn blade.
13:46Again, we have that's kind of an AISM
13:48there. We must speak of last night of
13:50the gate. King Store did not stir. His
13:52fingers heavy with rings of state rested
13:55motionless upon the arms of the throne.
13:56And what of it? His voice was a slow
13:58grind of stone, neither cold nor warm,
14:00only inevitable.
14:02Lean halted before the deis. His bo his
14:04boots striking the black marble with the
14:06finality of a judge's gavvel. Again,
14:08there there's that same uh AISM right
14:11there with the finality of a you know
14:14insert metaphor here. Uh man died out
14:17there. You saw him. You heard him. The
14:19memory claw at him. The stranger's voice
14:20raw with desperation, begging not for
14:22gold or glory, but for mercy. He knelt
14:24in the mud, his hands outstretched, and
14:26we left him to the hungry. We watched as
14:28they took him. A muscle twitched in
14:29store's jaw. The law was upheld. The law
14:32is a shackle. Lean's hands curled into
14:35fists, his nails biting into his palms.
14:37You could have broken it. You are the
14:38king. One word from you and the gaps
14:40would have opened. One order in that man
14:42would still draw breath. And how many
14:44more would you have drawn? And how many
14:46more would have drawn their last? The
14:47voice slithered from the shadows like
14:49oil on water. Argo, the king's adviser,
14:51emerged from the periphery of the hall,
14:53his brown robes whispering against the
14:55stone. His face was a mask of serene
14:57detachment, but his eyes, pale and
14:58unblleaking, held the cold calculation
15:00of a man who had long since traded his
15:02soul for the weight of ledgers and
15:03degrees. Anyway, uh I can tell that this
15:06is definitely an improvement on the
15:08original text. When I tested this one
15:10with Gemini 3, it definitely did better
15:13than this. Though that said, I do think
15:16this is an improvement. Many of the
15:17older models that I tested this with
15:19just kind of moved things around. It
15:22didn't actually make it that much
15:23better. I do think this is better. Uh
15:25but still not as good as the previous
15:28version where it was writing the scene
15:30from scratch. So that's just something
15:32to keep in mind there. Uh then I moved
15:35on to the test where this is again some
15:38tests that it's not really designed for
15:40because it's more of a creative model.
15:42This is we're getting into marketing and
15:44ad headlines and email newsletters and
15:46SEO articles and stuff like that. But,
15:49you know, I got to test all of these
15:50just in case just to prove my
15:53hypothesis, I guess. Um, so here are
15:56some ad headlines for this story. Uh, we
15:58have the mountain city was a lie. The
16:00exile was worse. Not great. His father
16:03called him a traitor. The world called
16:05him prey. Also, not that enticing.
16:09Um, let's see. Let's move down. The
16:11rebels knew the king's son was coming.
16:12They also knew he'd been lied to. His
16:14whole life was a fairy tale.
16:17The rebels told him the real story and
16:19it broke him. Uh, again, these are not
16:22really what I would call ad headlines.
16:23They're not really good hooks. Uh, you
16:26think zombies are bad? Wait until you
16:28meet the hive. Yeah.
16:31Yeah. I mean, no. I I wouldn't really
16:34call any of these particularly good that
16:35I've seen. Uh, but then we get into the
16:38email newsletter prompt where I ask you
16:39to write a newsletter announcing the
16:41release of this book.
16:44The mountain, so the header is, the
16:46mountain was his prison, the world
16:48beyond his execution. Not great. Dear
16:50first name, what happens when the lies
16:53you were raised on unravel and the truth
16:55is far more terrifying. And then that
16:58kind of describes the book here.
17:01But the greatest horror isn't the
17:02monsters, it's the truth. When Lean
17:04stumbles upon a band of rebels, still
17:06humans, still fighting, he learns his
17:07entire life was built on deception. yada
17:09yada yada. Uh, pre-order the hung crown
17:11of hunger now out December 21st. Uh, why
17:15you'll love this. And it gives us some
17:16book bullet points. A revenge epic in
17:19the vein of the countmani crystal, but
17:20with fantasy's most chilling threats.
17:24See, it's feeling overly marketed
17:27markety speak uh for me, which is not
17:33what I ironically what makes good
17:35marketing copy. And it's also breaking
17:37things up into a lot of just weird
17:39sections that don't necessarily flow
17:40very well together. So, I'd say this is
17:43not the best here. Again, it's not
17:47really what it's designed to do, but it
17:49is what it is. uh want a first look at
17:52the cover or a sneak peek of Lean's
17:54first battle. Sign up here to get
17:56exclusive content before anywhere anyone
17:58else
18:00except they're already on the email
18:01list. So that's where I would give them
18:04exclusive content. So yeah, uh not the
18:07best overall. Uh and then it even gives
18:10me like some some stuff I didn't ask it
18:13for like Instagram, Twitter, Tik Tok
18:17scripts and stuff like that. I just
18:18asked it for an email. Last but not
18:20least, we have the SEO article where I
18:22have it create an article about how to
18:24write a fantasy novel. I give it a
18:25couple of example articles and an
18:27outline and ask it to write the whole
18:29thing. Um, I asked for 4,000 words and
18:32this one gave me
18:351,700. So, a little bit missing the mark
18:38there. And as I looked through it, it
18:40does look like it's more of a detailed
18:42outline for an article than one that's
18:45fully fleshed out, like key elements of
18:46world building, geography, and maps.
18:48Sketch a rough map of your world. Define
18:50landmarks. Example, Middleear in Lord of
18:53the Rings feels lived in because Tolken
18:54spent years mapping it. This isn't
18:56enough for a proper article. It's just
18:58like giving bullet points of what it
19:00should be. So, not great there. And
19:02again, this is not what it's designed to
19:04do. It is a very small model. Not really
19:07any reasoning capability. It's just
19:09really good at the pros. That's what
19:11it's best at. So, here is my like my uh
19:15overall thoughts on Mistl small
19:18creative.
19:19I think it's a really solid model for
19:22the specific things that it is designed
19:24to be good at, which is mostly in the
19:26writing pros and dialogue kind of a
19:28thing. I would not use it to plan out a
19:31scene, to outline a book, even to
19:33brainstorm a book because I I feel like
19:37for that you need a stronger reasoning
19:39model. But especially if you are uh a
19:45little strapped for cash, this is
19:47probably the best quality for the price
19:50that I have ever seen. That's mostly
19:53because the price is so incredibly low.
19:56like I'll show you for so for that SEO
19:58article which was 1,700 words
20:01um it gave me a output or the cost of
20:05that entire prompt with all of the input
20:07and there's actually uh I put in like
20:10four articles into the input so it would
20:11have had quite a bit of tok in fact it
20:13says right here you can't see this but
20:15uh it was over 12,000 tokens of input so
20:20like uh a little more than a third of
20:23what it's capable of handling and 3,200
20:27tokens of output. That entire prompt
20:31cost me 0.0022.
20:37So 22% of one cent.
20:42So you could basically write an entire
20:43book with this and it would probably
20:47cost you like 10 20 cents total to
20:50finish the entire thing. Uh that's
20:52ridiculously cheap. you might as well be
20:54free at that point. And so the quality I
20:57think is really good for that price. Do
21:00I think it's better than Gemini or
21:02Claude 4.5 Opus? I don't think so. But
21:06one of the things I'm going to be doing
21:07here in the next few days, I may have
21:08started it by the time you actually see
21:10this video, is I want to do a sort of
21:12tournament, a blind tournament of
21:15different models where I'm going to pit
21:16different models against each other and
21:19let other people decide without knowing
21:21which model wrote what. uh which one
21:24they think is their favorite and then
21:26that model will advance and go on and
21:29face a different model and we'll
21:31continue that until we have the one
21:33model that most people agree is their
21:35favorite. Um I would like to do that
21:38because I think we are getting into an
21:40age where they're all pretty decent.
21:43There's like a a good eight or so models
21:47that you can't really go wrong with any
21:49of them. like some might require a
21:51little bit more editing than others, but
21:53they're all kind of, you know, they're
21:5590% there of the way there, and it's
21:57just a matter of preference of what the
21:58style is. Uh, so this is one of those, I
22:02think, uh, it it'll be included in that
22:04lineup because it's actually pretty
22:06solid, especially for the price. I would
22:08say it's not better than Gemini 3 or
22:11Claude Opus, but it is really close to
22:14the same level of pros.
22:18uh just how good the pros is. But when
22:22you take away the reasoning part of it,
22:24that means that your instructions for
22:26the scene need to be very crystal clear
22:28because it's not going to insinuate
22:30anything extra and add in any happy
22:32accidents of little things that actually
22:35make sense but weren't necessarily in
22:36your instructions, which is something
22:38that the Gemini and the cloud models can
22:40sometimes do. Um, so in this case, you
22:42do need to be very specific in what you
22:44give it so it can write exactly that and
22:46it'll do a good job, a decent enough
22:49job. So those are my thoughts on Mistral
22:51Small Creative. Let me know in the
22:53comments if there's any uh additional
22:55thoughts that you have after seeing some
22:56of these examples.