"Spikosama" is a hybrid model that combines a novelist-style AI with a rich vocabulary and syntax, and a chatbot/instruction-tuned AI that excels in logic and responsiveness to instructions.

If you plan to use "Spikosama", we recommend starting with the "New Assistant" option.

While you can use it by creating a document in the normal novel mode, please delete the initial text.

Users who mainly use the service for novel writing can continue writing their existing documents with "Spikosama". If you want it to simply continue writing without any instructions, press "Send" with nothing in the input field.

If you use a document that already contains User/Assistant blocks generated by "Spikosama" with older models like SuperTrin or Damsel+RAY, the AI may be influenced by the instruction syntax and stop writing long passages, or its behavior may change.

In this case, manually removing the user-inputted instruction syntax and the User/Assistant blocks (the icon parts in the image above) will restore the original behavior.

"Spikosama" is not compatible (will not function correctly) with the officially supported Chat Prompts or most existing scripts.

If you want to use "Spikosama" in a way similar to Chat Prompts, please create a document using "New Assistant".

"Spikosama" is a model that blends two personas: a chatbot AI and a novelist-style AI.

You can choose from three different modes, each with a different degree of instruction tuning, to suit your preferences and the situation (※The 【Solid】 mode cannot be selected in the Lite version). Here, we will explain the modes other than the default.

☆ Solid

This mode has a tuning closer to a general chatbot than the default, allowing it to write more logical sentences. However, due to its nature, it can sometimes become oddly argumentative or the sentence structure can become monotonous.

It is effective for prompts that require a high level of comprehension, such as when you absolutely want it to follow instructions, have it plan the next development, describe a complex scene, or play a pseudo-game.

♫ Colorful

When this mode is selected, output is generated by Damsel+RAY for Platinum members, and by SuperTrin for all other users.

These two models are specialized in generating natural novel continuations and are not designed to actively follow instructions, but they are superior to the default mode in writing natural conversations and narrative text, and in advancing the scene one after another.

A useful technique is to first have the default or Solid mode write the beginning of each block, and then use Colorful mode to continue writing the rest.

Editable Anywhere

Unlike typical chatbots where the user's input and the AI's output are clearly separated, "Spikosama" allows you to edit the AI's output afterward. Also, instead of outputting the entire response in one block, it generates text incrementally, similar to other AI Novelist models.

If the output text ends midway, or if you want it to write a longer continuation, press the "Send" button without typing anything in the input field.

Making it Continue Writing

Models that support instruction tuning, such as "Spikosama", are designed to stop generating at a good breaking point for each part.
If you see a message like Please say something to continue., or if it almost stops writing, you can get it to continue writing indefinitely like older models by typing an instruction like "Continue." or "Please continue" in the chat field.

This might be confusing for those used to older models, but once you get used to it, you'll find it convenient as it prevents the scene from changing without your consent.

※ You can also make it write continuously like older models by unchecking "Stop output at each paragraph". This action does not use the thinking block.

Let's Edit It

Unlike other chatbot-type AIs with a fixed writing style, "Spikosama" is tuned to write based on the format and style of the text written so far, similar to previous AI Novelist models (especially the Default mode is strongly influenced by the output so far, while the Solid mode tends to regularize the style to some extent).

If there are parts of the AI's output you don't like or feel are "not right," don't leave them as they are. Directly rewriting or partially deleting them will make it easier to generate the kind of text you want.

Interrupting with Instructions

You can also send a new instruction syntax in the middle of the AI's output. This is effective for forcibly changing the scene or viewpoint, or for making it analyze the situation.

About Advanced Options

In "Spikosama", the generation parameters in the advanced options are fixed and automatically adjusted, except for the items listed below (values other than these will not be reflected except in "Colorful" mode).

・ Randomness (in the range of 0.95 to 1.50)
・ Repetition Penalty
・ Typical P
・ Output Length
・ Context Length for AI to Read
・ Character Book Priority
・ Character Count to Scan for Character Book
・ Formatting

Handling of Footnotes

In a document containing User/Assistant blocks, the footnote is inserted at the beginning of the last (bottom-most) user instruction line in the text.

Example:


あなたは夏目漱石の小説に出てくるネコのワガハイです。


ここに脚注が挿入される
ワガハイが書生について語るシーンを描写。

Generating Text from a Plot

Unlike general-purpose chatbot AIs, "Spikosama" has had its dataset created and tuned with a strong focus on generating body text from bullet-point plots. When doing this, be mindful of the following two points:

・ Enter the elements you want to be described in the next 300-1000 characters.
If you send a prompt with too many elements at once, it may not be reflected well or may become verbose.

・ Don't send the entire plot structure (beginning, development, turn, conclusion) at once.
If you feed the entire plot at once, even scenes you want described in detail will end quickly. It's recommended to write and send one, or at most two, upcoming events.

Making it Fill in the Next Instruction

If the text ends with a user icon or [#User], pressing the send button without writing anything will make the AI fill in the rest of the instruction prompt for you.

Example:

Behavior of Repetition Penalty

For the instruction-tuned model "Spikosama," the actual behavior of the repetition penalty has a marked difference from previous models. Here is a summary of simple examples.

↙ When Repetition Penalty is Low (up to ~1.05)

◆ Default:

・Text follows instructions: Good
・Spontaneous new developments: Fair
・Maintains syntax/format, slightly monotonous text

☆ Solid:

・Text follows instructions: Excellent
・Spontaneous new developments: Very Poor
・Maintains syntax/format, monotonous text

♫ Colorful:

・Largely ignores instructions
・Spontaneous new developments: Fair
・Maintains syntax/format

↗ When Repetition Penalty is High (1.20 and up)

◆ Default:

・Text follows instructions: Fair
・Spontaneous new developments: Good
・Tries to break syntax/format

☆ Solid:

・Text follows instructions: Good
・Spontaneous new developments: Fair
・Varied syntax

♫ Colorful:

・Largely ignores instructions
・Spontaneous new developments: Excellent
・Tries to break syntax/format

A high repetition penalty creates a force that tries to break the syntax that has been continuing in the text and do something new. For example, when there is a specific format, such as in an RPG battle simulation, we recommend setting the repetition penalty to a lower value if the text has a specific format.

About Thinking Models

"Spikosama-02" and "Spikosama Max" are what are known as thinking/reasoning type models. When outputting instruction-tuned text, they first use a thinking block to verify and plan the prompt, thereby achieving both high creativity and prompt adherence (the degree to which it follows instructions).

The thinking block can only be used in Default and Solid modes. Colorful does not use a thinking block and outputs as before.

"Spikosama Max" has been specifically post-trained with multi-turn, long-context thinking data. For this reason, it may think longer when the situation is complex or when complex instructions are given.

Limitations

Only the most recent thinking block is sent to the AI (this is in line with many other thinking models).

To use the thinking block, you must send some kind of instruction to the AI using the chat field. Simply typing "Please continue" is sufficient.

No-think/nothink Models

For "Spikosama-02", we have prepared a variation model that skips the thinking block. Please use this when you want to shorten the thinking time, or when you want to finely adjust the output content with scripts.
※Please understand that performance will drop somewhat in situations requiring complex logic or for long-context texts.

Actual Example

An example of how the model actually thinks and outputs is shown below.

Example:


吾輩は猫である。そんな吾輩の最近の発見と癒やしは……


<think>
この文章は、夏目漱石の名作『吾輩は猫である』の書き出しを引用しています。「吾輩は猫である」で始まるので、これは明らかに『吾輩は猫である』の文体や口調を真似た文章だと言えます。

続きを書くにあたって、このスタイルを維持しつつ、「最近の発見」について書いていく必要があります。漱石の文体の特徴としては：
1. 猫の一人称視点からの観察
2. 人間社会への冷静な批評や皮肉
3. 比較的長めの文章構造
4. わずかにユーモアを含む分析的語り

これらの特徴を維持しながら、現代的な要素や著者の独自性も加えた続きを書いていきます。「最近の発見」という言葉から、何か新しい発見や気づきがあったことが伺えます。それが何なのかを展開していきましょう。

この文体の特徴を生かしながら、現代社会での猫としての観察や発見について書いていきます。特に、デジタル社会における人間の行動や、現代の環境についての観察などが適しているでしょう。

また、「吾輩は猫である」の文体はやや古風で、時に哲学的かつ批評的な視点を持っているので、その要素も取り入れていきます。
</think>

# 吾輩は猫である。最近の発見について

吾輩は猫である。名前はまだない。いや、正確に言えば名前はあるのだが、吾輩は覚えぬ。人間どもは勝手に「タマ」「クロ」「ミケ」などと呼ぶが、その名前で呼ばれたところで、吾輩の毛色や瞳の色が変わるわけではない。

さて、最近の吾輩の発見であるが、この時代の人間どもは「スマホ」という小さな箱に異様に執着している。四六時中その小さな画面に見入っている姿は、まるで吾輩の生前の主人、苦沙弥先生が煙草を嗜んでいた頃を思い起こさせる。違うのは、あの頃の先生は何かを吸うために煙草を持っていたが、現代の人間どもは画面に向かって何かを「吸い込む」わけでもなく、ただただ指でその画面を擦り、押してはいるだけだ。

最近ではそのスマホという物から奇妙な音が聞こえてくることも多くなった。時折、「ピーン」と鳴ったり、「ピロリロ」と鳴ったりする。吾輩がその音に耳を傾けると、人間どもは慌ててスマホを手に取り、その小さな画面に向かって何やら指を走らせる。

これが吾輩の最近の発見である。人間どもはこの「ピーン」や「ピロリロ」の音に異常に敏感なのだ。まるで吾輩が夜中に突如として「ニャー」と鳴くと人間どもが慌てて起きるようなものである。しかし彼らが起きたところで、吾輩は何も求めているわけではないことが多いのだが。

さらに面白いのは、彼らがそのスマホに向かって話しかける姿である。吾輩は時折、「吾輩に話しかけているのか？」と思うこともあるが、どうやら空に向かって話しかけているらしい。あるいは「Siri」と呼ばれる何かと会話をしているようだ。

苦沙弥先生の時代には考えられない光景である。あの頃は「電話」という大きな機械があったが、それでも人間同士が会話をするために使っていた。しかし現代ではその電話が小さく、しかも誰もいない場所に向かって話しかけているのである。

吾輩は思う。人間とは奇妙な生き物である。そして吾輩は猫である。そんな吾輩は今日も窓辺で人間どもの姿を眺めながら、「ピーン」や「ピロリロ」の音に耳を傾けているのである。

JA-JP Benchmark Score

Model	JSQuAD 0-shot	JCommonsenseQA 0-shot	MMLU-JA-JP (OpenAI)
Spiko-Max 202506 (thinking budget: 1536 tokens)	0.956	0.975	0.868
Spiko-02 Solid (thinking budget: 1024 tokens)	0.950	0.980	0.853
Spiko-Max 202503 (thinking budget: 1024 tokens)	0.942	0.979	0.838
Spiko-01 Solid	0.847	0.911	0.761
Damsel+RAY (dev)	0.78	0.93	-
SuperTrin α2	0.71	0.70	-
*o3-high	-	-	0.890
*o1	-	-	0.888
*gpt-4.5-preview-2025-02-27	-	-	0.869
*gpt-4.1-2025-04-14	-	-	0.856
*gpt-4o-2024-11-20	-	-	0.835
*o3-mini	-	-	0.822
*gpt-4.1-mini-2025-04-14	-	-	0.810
**Llama3.1 405B	-	-	0.80
**mistral-large-2407	-	-	0.78
**Llama3.1 70B	-	-	0.72
*gpt-4.1-nano-2025-04-14	-	-	0.690
**Command R+	-	-	0.67

* Reported by OpenAI
** Reported by Mistral AI