Model Recommendations:

Pre-roll:

Hey guys! Hopefully I don't take up too much of your time. I'm working on making a modern model recommendation page since a lot of us (me included) struggle to pick out new models and understand things such as: Why Wizard over Darkplanet? What's the difference between Mistral vs. Llama? I could really use your input on what features are most important to consider when choosing a new model. In other words, because I don't have the time, wherewithal, or knowledge to test these models in every meaningful fashion, I was hoping to crowdsource it with the community and hopefully make this page as helpful as possible for all of us.

Right now I'm asking for three things:

A. If you have any info that would help in filling out the table.

B. If you'd be willing to test any of the listed models with your favored cards and come back with what preset you found to work best, what cards the model liked, what strengths & weaknesses you've found as well as any info that is listed in the table (Context Length | Roleplay Style | Best For | Bad At | Local model (Yes/No/Not Recommended) | Additional notes?)

C. If you have any further model recommendations and why.

D. What models you know are good at very specific things? I.e being able to write wlw, mlm, demihumans, etc.

Model Recommendations Table

Legend:

• HF - Hugging face
• F - Featherless
• A - AlriAI
• I - Infermatic
• O - Openrouter
• C - Cloud
• L - Local

Reviews & Raw Data:

1. Wizard 8x22b is Mistral v3 | 65536 (effective up to about 30k-ish) | (at least) HF, O, L | It's on the smarter and descriptive side, but has issues with being creative and it writes too much. | Cheap, pretty smart, uncensored | Can feel bland and write too much. | Not Recommended | HF & O | I would consider this model to be a little bit outdated; it's not free unlike Mistral Large, for example, but it's less smart and I personally consider it very boring. It can be tuned a bit with prompts, but it's still an assistant model through and through
2. WizardLM-2 8x22B
   Model Type: Mistral Mixtral 8x22B
   Context Length: 65,536 (O), 16K (F & I)
   Available on: F, I, O
   Roleplay Style: Any. RPG, SFW, ERP
   Best For: SFW and RP with heavy lore, Hurt/Comfort
   Bad At: Dark RP (due to heavy positive bias)
   Creator: Microsoft
   Local?: Not Recommended
   HF: https://huggingface.co/alpindale/WizardLM-2-8x22B (This is a repost, original is taken down by Microsoft and never reposted)
   Featherless: https://featherless.ai/models/alpindale/WizardLM-2-8x22B
   Additional notes: Smart and very good spatial awareness. Good at creative writing. Suffers from GPT slop and strong positive bias, but can still act unhinged

---

4. Zyph:
   Honestly, I dont really have all the technical LLM stuff down so much as I go by feel. my go to models are:
   

• Anthracite Magnum 72b v4 - because I just in general think its a great multipurpose model that keeps track of a story pretty well and tends to not get too stuck on an odd detail here and there like a lot of the other ones Ive tested. It tends to be a little... thirsty though. So I sometimes edit out the NSFW portion of responses and continue on and it seems to keep pretty SFW with a little handholding
• Euryale 70b 2.3 - speaking of handholding... Euryale is THIRSTY sometimes. like, way too much unless you're specifically looking for NSFW roleplay. HOWEVER, with some handholding and keeping things PG 13, the responses it gives are fantastic and emotion-driven. This is my go to model for drama roleplay, especially angsty drama because it does such a great job. Qwen 2.5 72b Kunou is supposed to be even slightly better and trained on a Qwen base like Magnum 72b is, so Im eager to give it a go, but havent gotten around to it yet.
• Deepseek - Its just fun to play around with, and once you have a decent preset to rein it in it does pretty well. Where it really shines, though, is fantasy adventure roleplay. This is the first model that played one of my characters they way I intended them to be played rather than very surface level. For Mallory, other models would just be like, "oh, she uses daggers, shes a rogue and she rushes you with aflurry of attacks" but deepseek caught on to her use of ice magic and illusion magic as well, and without me strong arming it into using it, got super creative and was using cool combos and everything. it really turned her into a powerhouse to fight against and it was super awesome! it can do drama RP as well, of course, but for fantasy adventure, its the best I've EVER used.

---

5. Check reddit.
   https://www.reddit.com/r/SillyTavernAI/s/Xfp7s1fIlS

---

6. Model chat in Zyph disocrd March 30th
   Anubis Pro 105b sounds good, it's from a good model maker albeit drummer models are usually a bit horny (llama 3 instruct settings)

Electra R1 is very good (llama 3 instruct settings)

I don't use OR very much tbh
(and, of course, use a good prompt and post history for them.)
But when I can afford it? Claude sonnet. The problem though is modifying your prompting. If you play nice with Claude though, it's some of the best there is to be had for roleplay.
I've... Never seen it be repetitive.

But the main thing is giving it a valid reason to get erotic.

In my experience there had to be a positive emotion tied to the act. It can't be just... Smut. Then it pushes back and doesn't wanna play anymore.Another big issue with Claude is when your context gets high it'll bleed through money really fast

I stand by Magnum 72b for being a good multipurpose general model. Euryale 70b is my favorite for emotionally intelligent, angsty RP. Both of them can be a bit on the horny side, so depending on how defs are written and the setup, need some handholding at first to steer clear of NSFW, but on the whole responses are great. I’m having a lot of fun with Deepseek right now too though

.ive been revolving around (on featherless) using deepseek v3 (not a great rp model, but can be fun), Electra R1, Nevoria R1, Genetic Lemonade Final, and locally Stardust V2. Mistral small also has a really good rep
command A.Cohere delivers solid models for regular RP but some less-friendly stuff is really, really hard for it.I tried Command R on my murderous torture elf and it would all threaten and imply but nothing would happen.Meanwhile, Mistral would go Casino Royale on deez nuts.
Oh by the way, right now I'm rolling with fallen gemma3
and it litereally killed me on elf slave like she just randomly slid my throat with piece of glass. it supposed to be a savior card, Its insanely good for 27b

but Fallen Gemma 3? [Evil tine of Gemma 3 27B - knock out positivity.] well, no, that was not horny
that was unhinged
its just not really holding back
but if you give a hint like
"my towel ride a little lower on my hip"
its just picking up immidiately, and its first model, that I enjoy playing open RPG cards with

---

7. Oh boy, it's been a while since I used locals other than Hanami (kek)
   For 70b I think people have been singing praises for merges from one guy, Steel (who also merged Nevoria)
   That includes his newest model Electra and his three sisters: Cu-Mai, San-Mai and Mokume Gane
   I haven't experienced them but you can tell how popular they are based on their leaderboard:
   https://www.arliai.com/models-ranking

For me, other models I still go back to:
Chuluun v0.08 - A Qwen merge that miraculously works well with high temp (1.2 - 1.4), very good prose and is still the most popular Qwen model on Arli. It has ink as merge, which de-slops most common slops and I've noticed the same when I tested it

Llama-3.3+(3v3.3)-70B-TenyxChat-DaybreakStorywriter
This is that Tenyx DB from infer, now on Arli with llama 3.3 instruct as base. It actually feels nice and is approved by the former infer Tenyx gang, maintains the novel writer prose without getting too horny

MiraiFanfare (& v2) - this is one of the earlier merges for llama 3.3 - a very cursed one. Seriously, it's takes a cursed merge with a bunch of east Asian languages, Nemotron and Euryale, then slaps it on top of Eva, but it gives me the most unique prose. Short punchy sentences and insane creativity. The downside will be that it sometimes writes very short (2 - 3 paragraphs)

---

Big model creators:

Primary:

• Microsoft
• Google
• Laititude (Hermes, Wayfarer)

Secondary:

• TheDrummer (Anubis, Behemoth, Evil Gemma)

Starter Preset's:

Recommended System Prompts

Good to Know

Base Model Pointers

Prospective Other Models

1. https://huggingface.co/sphiratrioth666/SX-2_Characters_Environment_SillyTavern
2. https://featherless.ai/models/Steelskull/L3.3-Mokume-Gane-R1-70b-v1.1/readme
3. https://huggingface.co/Black-Ink-Guild/Pernicious_Prophecy_70B
4. https://huggingface.co/TheBloke/Chronomaid-Storytelling-13B-GGUF

Gratitude Goes Out To:

• Pepper, for inspiring me with their original table, and their amazing updated one that I used as a major source
• Feldherren, for pushing me to start this

Contributers:

StatouTW, Feldherren, BexRU, OMGWTFBBQ, zypher,

My Main Page