> For the complete documentation index, see [llms.txt](https://franofran.gitbook.io/franofran-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://franofran.gitbook.io/franofran-docs/guides/proxies-options.md).

# Proxies Options

<mark style="color:yellow;">**Trying to make this guide ADHD and dyslexia friendly, feel free to let me know if I can improve the presentation for that!**</mark>

Here is a simplified version, read below for more details!

<figure><img src="/files/qFXiEYSZDOZptGyjZBWW" alt=""><figcaption></figcaption></figure>

## <mark style="color:orange;">Basic AI Lingo:</mark>

* <mark style="color:orange;">**Tokens**</mark> = A token is a unit of text used by AI models. A single word can be 1-3 tokens depending on its length and complexity.

* <mark style="color:orange;">**Memory Context**</mark>= How much the AI remembers of your RP, measured in Tokens. Always includes the bot permanent tokens and your persona's, any remaining free space will use temporary tokens(first message, dialog examples and Rp chatlog). the longer you go on the chat the older temporary tokens will be forgotten)

* <mark style="color:orange;">**B**</mark>= Usually shows on a model in front of a number, it basically means how smart a model is, usually they go like 12B, 32B, 70B etc. The higher the number the smarter the model is.

***

## What I look for in an AI model(coming from the JLLM):

* <mark style="color:orange;">**Good memory**</mark>\
  Jllm only has 3k-7k context nowadays, times of high traffic like Christmas lowers it a lot. I usually look for something with 16K+ of context.
* <mark style="color:orange;">**Avoiding cliche behaviors and phrases**</mark>\
  Examples: "maybe...just maybe", "jolts of electricity", "I'll ruin you for everyone else", etc., you know what I mean
* <mark style="color:orange;">**Avoid overly horny behavior triggered for nothing**</mark>\
  So you can have slow burn Rps, and/or Rps with plots that don't focus on NSFW.
* <mark style="color:orange;">**More intelligence, be more in character, more knowledgeable**</mark>\
  ( for my case, more knowledgeable about the One piece lore).
* <mark style="color:orange;">**Better special awareness + more logic**</mark>

###

## <mark style="color:orange;">**FREE PROXIES OPTIONS**</mark><mark style="color:orange;">:</mark>

***

### <mark style="color:green;">Kobold proxy collab</mark>

<mark style="color:orange;">**Pros:**</mark>

> * Has a lot of variety of small models from 7 to 22B, mostly 12 B's
> * 24K memory context on 12B models even 16K on 14Bs
> * Fast generation
> * Medium easy integration on jai

<mark style="color:orange;">**Cons:**</mark>

> * A hassle to setup: Every time you want to use any chat, you have to manually set up the url at the proxies settings. Additionally, whenever you want to activate the proxy, it takes about about 5 min to boot it, then you have to set it up and only then you can chat. Also, if you close the collab browser page, it will kill your API key and you'll have to set it up again(wait the 5 min loading time again) so you can use it again. I don't recommend this if your wifi fails a lot

> * I'm not sure on this one, but I heard there is a time limit for every day, I think 3 hrs, but I don't remember. I didn't use it much due to the hassle to set up

<sub>**Feel free to explore, but these are just some cool model recommendations, I enjoyed!**</sub><sub>:</sub>

* <sub>EVA-Tissint(14b) | personal fav from there</sub>
* <sub>Starcannon(12b) | Personally, I liked it</sub>
* <sub>Magnum(12b) | people say it's good</sub>
* <sub>Mag-Mell(12b) | people say it's good</sub>

[Here is the google collab link(done by Hibikiass)](https://colab.research.google.com/drive/1l_wRGeD-LnRl3VtZHDc7epW_XW0nJvew#scrollTo=pf4AQOYgTB2d)

***

### <mark style="color:green;">Kobold Local Host</mark>

You can run your own llm on your own pc

<mark style="color:orange;">**Pros:**</mark>

> * You don't have to share the machine with anyone, I heard this gives you better response even with the same model
> * You can choose any model you want and it will be free, there are thousands of options

<mark style="color:orange;">**Cons:**</mark>

> * Hard to install and setup at first
> * Like the Kobold Collab, You'll have to boot it every time you want to use it
> * You need a powerful PC with at least 6GB of VRAM to run even a small 7B model.

***

### <mark style="color:green;">Arli AI</mark> <a href="#arli-ai" id="arli-ai"></a>

They have a lot of models under their 27B genma line!

<mark style="color:orange;">**Pros:**</mark>

> * Easy to set up. Once you have it set up, you don't need to do anything else no more. Plug and play experience!

<mark style="color:orange;">**Cons:**</mark>

> * Slow response times. In my previous experience, I was as waiting 30s to 1 min per generation before the first word was generated.
> * I liked the models, but I still felt like it lacked what I looked for for long term/complex RP, especially when it came to have knowledge of canon characters, however this bullet point is totally personal preference and you might have a totally different experience!

[Link to Arli's web page](https://www.arliai.com/)

***

### <mark style="color:green;">Open Router</mark>

<mark style="color:orange;">**Pros:**</mark>

> * Has some different free models, Most of them are not entirely made for RP, they do a pretty good job!
> * Most models have an extremely huge context
> * Easy to set up. Once you do it, you don't have to touch it anymore

<mark style="color:orange;">**Cons:**</mark>

> * A lot of the free models aren't tailored to RP, so you will probably have to make a custom prompt to tailor it to your taste and play with the temp(at least from my experience)
>
> * Some lots of errors for free users since they prioritize the paid users

<mark style="color:orange;">**Model recs**</mark><mark style="color:orange;">:</mark>

* <sub>Deepseek R1 (free) | Read more about it below at the paid Open router section.</sub>
* <sub>Deepseek V3 (free) | Read more about it below at the paid Open router section.</sub>

[Link To open router](https://openrouter.ai/)

***

##

## <mark style="color:orange;">PAID PROXIES OPTIONS</mark>

There are usually 2 general options within paid proxies. Subscriptions or pay per token(basically you paid as much as you use)

### <mark style="color:orange;">SUBSCRIPTIONS</mark>

### <mark style="color:green;">Arli ai:</mark> <a href="#arli-ai_1" id="arli-ai_1"></a>

<sub><mark style="color:orange;">**Pricing**<mark style="color:orange;"></sub><sub><mark style="color:orange;">:<mark style="color:orange;"></sub>\
10$/m = `Access Up to 72B Models(16k context MAX)`\
15$/m = `all models up to 235B(32k context MAX)`

<mark style="color:orange;">**Pros:**</mark>

> * Unlimited messages
> * Multiple diverse models
> * They keep adding more models!
> * Easy to set up. Once it's done you don't need to touch it anymore unless you want to change models. Plug and play experience.

<mark style="color:orange;">**Cons:**</mark>

> * Slow replies, expect 30 to 1 min for the message to start generating the first word

<sub>Personal fav models from what I tried with my bots as of January 2025:</sub>

* <sub>Llama-3.3-70B-DeepSeek-R1-Distill</sub>
* <sub>Qwen2.5-72B-Evathene-v1.3</sub>
* <sub>Llama-3.3+3.1-70B-Euryale-v2.2</sub>
* <sub>Qwen2.5-72B-EVA-v0.2</sub>

***

### <mark style="color:green;">Other subs I'm aware they exist but I didn't try</mark> <a href="#other-subs-im-aware-they-exist-but-i-didnt-try" id="other-subs-im-aware-they-exist-but-i-didnt-try"></a>

<mark style="color:orange;">**Infermatic**</mark><mark style="color:orange;">:</mark> [Link](https://infermatic.ai/pricing/)

<mark style="color:orange;">**Featherless**</mark><mark style="color:orange;">:</mark> [Link](https://featherless.ai/#pricing)<br>

***

### &#x20;<a href="#pay-per-token" id="pay-per-token"></a>

### &#x20;<a href="#pay-per-token" id="pay-per-token"></a>

### <mark style="color:orange;">PAY PER TOKEN:</mark> <a href="#pay-per-token" id="pay-per-token"></a>

You pay as much as you use. Price depends on each model

### <mark style="color:green;">Open router</mark> <a href="#open-router_1" id="open-router_1"></a>

[Link](https://openrouter.ai/)

<mark style="color:orange;">**Pros:**</mark>

> * Extreme variety of models with extreme variety of price ranges, intelligence, most have huge context memory

<mark style="color:orange;">**Cons:**</mark>

> * If you are a reroll maniac, you can end up with a huge bill, fortunately you can put a limit to how much I wanna spend!

<sub>Personal fav models:</sub>

* <sub>DeepSeek V3 0324</sub>
* <sub>DeepSeek R1</sub>
* <sub>Nemotron Ultra</sub>
* <sub>Gemini 2.5  Pro</sub>

***

### <mark style="color:green;">DeepSeek</mark>

You can Choose in between the V3 and R1 (read about them at the open router section!) Their models are a lot more stable here than on Open router and cheaper(normal price but you don't compete with other provider's overpriced prices)

<mark style="color:orange;">**Pros:**</mark>

> * Lowest R1 or V3 prices! no errors like in open router even on the paid version. Some hours offer 75% of discount on R1!(good for Europeans)>
> * cheap considering other models
> * Faster  and more stable than on Open router

<mark style="color:orange;">**Cons:**</mark>

> * More or Same issues as the models on OR, as the issues are from the models themselves, However, I THINK the models perform better here and are more stable

[DeepSeek Platform](https://platform.deepseek.com/)

[Guide on how to Set up deekseek](/franofran-docs/how-to-setup-deepseek-official.md)

***

##

## <mark style="color:orange;">MODEL OVERVIEW PROS AND CONS</mark>

<mark style="color:green;">**This is my personal model overview. Other people might have different experiences and opinions!**</mark>

Prices <mark style="color:orange;">"Cheap" or "Expensive"</mark> are in comparison to having a subscription of about 10$/m

* <mark style="color:orange;">**Flow**</mark> = RP adaptability. How well it flows from message to message with a natural feel
* <mark style="color:orange;">**Aggressiveness**</mark> = How cruel/harsh a model is, tendency for cruelty
* <mark style="color:orange;">**Chaoticness**</mark> = How creative or out of rails a model is likely to be, more stars mean more chaotic
* <mark style="color:orange;">**Intelligence**</mark> = How good at understanding subtext or logic

<table><thead><tr><th width="97.800048828125">Model name</th><th width="100.7332763671875">Price</th><th width="107.93328857421875" data-type="rating" data-max="5">Intel</th><th width="98.5999755859375" data-type="rating" data-max="5">Aggro</th><th width="106.533203125" data-type="rating" data-max="5">Flow</th><th width="105.5999755859375" data-type="rating" data-max="5">Chaos</th><th width="106.0667724609375" data-type="rating" data-max="5">In Character</th></tr></thead><tbody><tr><td><mark style="color:orange;">JLLM</mark></td><td>FREE</td><td>2</td><td>3</td><td>4</td><td>2</td><td>2</td></tr><tr><td><mark style="color:orange;">DeepSeek R1</mark></td><td>FREE or cheap</td><td>4</td><td>4</td><td>2</td><td>5</td><td>5</td></tr><tr><td><mark style="color:orange;">Deepseek V3 0324</mark></td><td>FREE or Very cheap</td><td>4</td><td>3</td><td>4</td><td>4</td><td>4</td></tr><tr><td><mark style="color:orange;">Claude Sonnet</mark></td><td>Expensive</td><td>5</td><td>3</td><td>5</td><td>3</td><td>4</td></tr><tr><td><mark style="color:orange;">Nemotron Ultra</mark></td><td>FREE</td><td>4</td><td>2</td><td>5</td><td>2</td><td>4</td></tr><tr><td><mark style="color:orange;">Gemini 2.5 pro</mark></td><td>FREE or Expensive</td><td>5</td><td>3</td><td>5</td><td>2</td><td>5</td></tr></tbody></table>

### <mark style="color:orange;">**Claude sonnet 3.5 | Best quality | Expensive**</mark> <a href="#claude-sonnet-35-self-moderated-best-quality-expensive" id="claude-sonnet-35-self-moderated-best-quality-expensive"></a>

<sub>Available on Open Router and via their official API</sub>

<mark style="color:orange;">**Pros:**</mark>

> * Sonnet is king
> * Has a lot of general lore for canon characters
> * Characters feel in character a lot
> * Stable
> * Good plot and creativity
> * Best one regarding not being repetitive

<mark style="color:orange;">**Cons:**</mark>

> * Expensive if you use Jai a lot. Using it with 16k of memory context, messages can go from 1\~6 cents per message(from my experience), only use this on the regular if ur an oil prince/ess!
> * Censored. There are ways around it with prompts and prefill but if you want to use it uncensored, you need to boot the collab page every time and run it every time you want to chat(like in kobolt collab)

Extra notes: Wanna go bankrupt? You can also use Claude Opus, I heard it's really good!

***

### <mark style="color:orange;">DeepSeek R1 | Good For Price</mark>

<sub>Available on Open Router, Chutes and DeepSeek's Platform</sub>

<mark style="color:orange;">**Pros:**</mark>

> * Characters feel in character very accurately
> * Good body language and Mannerisms
> * Good spacial awareness, really good brain, especially with powers that require some logic and some LLms struggle to grasp the concept
> * As good as sonnet but for a much much smaller fraction of the price(in my personal experience, but needs a good custom advanced prompt)
> * A lot of canon character lore
> * Uncensored, no need for prefils/jailbreaks

<mark style="color:orange;">**Cons:**</mark>

> * Slow to generate(because it's a reasoning model, hence it's smartness). Faster than Arli's(from the time when I tested arli)
> * Issues with text formatting, but it's possible to mostly get rid of it with a good prompt, however its hard to get it consistent
> * Struggles with character development. It it will stick to characters too much not allowing much room for different behaviors, this can be improved with prompts, but it's still a high maintenance model that needs to be hand holded. It makes characters feel static reacting almost always in the same way
> * Not sure if it's a con, but it's really aggressive compared to other LLMS and WILL be mean and cruel if needed sometimes without reason

***

### <mark style="color:orange;">Deepseek V3 0324 | Best quality for Price</mark>

<sub>Available on Open Router, Chutes and DeepSeek's Platform</sub>

<mark style="color:orange;">**Pros:**</mark>

> * Good at being in character
> * Good price
> * Good character development and good prose
> * Wide canon characters knowledge

<mark style="color:orange;">**Cons:**</mark>

> * While it's uncensored, it does feel like anything that would typically be censored happens off screen and it doesn't focus on anything graphics like NSFW o violence, however if you have a good prompt you can bypass that
> * I think it has a little of positive bias
> * It feels insane with harsh topics, like its loosing it mind in a lot of cases

***

### <mark style="color:orange;">Nemotron Ultra</mark>

<sub>Available on Open Router</sub>

<mark style="color:orange;">**Pros:**</mark>

> * Has a good flow from text to text
> * FREE
> * Feels smarter than other models in smaller things, like understanding what it means being unconscious, understanding that in a multi bot, some characters don't participate in the response and such things
> * Feels sane compared to deepseek models
> * Adheres consistently really well to advanced prompts without "fighting with it"
> * You can trigger reasoning with OOC, making it analyze better the situation

<mark style="color:orange;">**Cons:**</mark>

> * Despite stable, it feels a little lacking in chaos/creativity

<mark style="color:orange;">**Notes:**</mark>

> * You will need a prompt to enable NSFW stuff, but one you have it there are no drawbacks