A Primer on Model Hosts:

For when you want something better then the free model but are too much of a caveman to set up local.

Skelly from AWR here again! Today's topic: Model Host platforms. Before we begin, I'd like to make some things clear. When it comes to technical details and tech in general, I am a caveman bashing rocks together in a world full of wizards who can cast spells of code. In addition, my involvement with the Wyvernchat community has left me with some bias towards Featherless, which can influence the final product despite my best attempts. As such, YMMV, and my advice is not anything near the be-all-end-all of discussion.

Introduction:

TLDR: If you have the money, I'd pick up Featherless for sheer choice and Deepseek. If you don't have cash, use OpenRouter's free models. If you feel overwhelmed by Featherless or want something less expensive that still has the big models, pick up ArliAI. If you want a specific older model or storytelling model, go to Infermatic.

OpenRouter: The Community's First Option

The largest and oldest host by far, OpenRouter is the go-to model host for individual and enterprise-level consumers.

Website: https://openrouter.ai/
Discord: https://discord.com/invite/fVyRaUDgxW
Cost Structure: PPT (Pay per token)
Most Popular Models:

  1. MythoMax 13B (One of the highest performing and most popular fine-tunes of Llama 2 13B, with rich descriptions and roleplay.) - 4096 Token Context
  2. WizardLM-2 8x22B - 65536 Tokens Context
  3. Deepseek V3/R1* (Censored) - 128K Context
Pros: Cons:
Hosts many of the community's most popular and well-documented models Pricing works of a cost-per-token basis, often cheaper for non-power users, but for roleplays that require multiple rerolls and testing, it can rack up costs quickly
A wide selection of models, rivaling Featherless in quantity and often including cutting-edge or specialized models Hosts of models aren't always running unquantified models, so performance might be slightly lower then local
Due to age, Most chatbot websites have some integration with the OpenRouter platform Harder to get in contact with staff due to size
Offers many of the most major popular models entirely for free (i.e. Deepseek & Mythomax; a classic model for free ) Models on OpenRouter are often censored, requiring Jailbreaks
Easiest to use UI Models can stop being free at anytime

ArliAI: A Promising New Competitor On The Block

The newest kid on the block, ArliAI, has been making waves with their service, delivering high-quality results with low input costs. Servicing the middle of the road, it is a promising player, with a responsive developer at its helm and unique offerings, though it's still working through growing pains and stability issues.

Website: https://www.arliai.com/
Discord: https://discord.com/invite/t75KbPgwhk
Cost Structure: Subscription—Multiple tiers (10 USD → 250 USD)
Quick use Guide: https://www.arliai.com/roleplay
Most Popular Models:

  1. QQwen2.5-72B-Chuluun-v0.08 - 24576 Tokens Context.
  2. ChatLlama-3.3+(3.1v3.3)-70B-Hanami-x1 - 24576 Tokens Context.
  3. ChatLlama-3.3-70B-Prikol-v0.5 - 24576 Tokens Context
Pros Cons:
Log free, uncensored and personally run by one man; truly the american dream Frequent slow periods and occasional outages due to server load
Offers DRY/XRC Samplers with overall excellent Sampler Support Integration issues with some frontends, though usually resolved quickly
Dev has been a part of the community for some time and develops models for open use for the community Limited model selection, primarily Llama and Qwen variants. (However, the developer is more then willing to add models upon request.)
No rate limiting; only parallel request limits (adjustable based on tier) While only for short periods thanks to an active team, the platform can be unstable
While more expensive then Infermatic, ArliAI offers a wider selection of 72B (and smaller) models However, expect that these will be severely slowed down or offline at peak hours

Infermatic: The Creators Model Host

The most budget-friendly host platform on this list. They focus on the quality of each individual model rather than offering a wide variety and overwhelming users. While offering some unique models and large context windows, the platform has faced major criticism regarding declining model quality and moderation.

Website: https://infermatic.ai/
Discord: https://discord.gg/9GUXmDx9GF
Cost Structure: Subscription (9 USD/20 USD)
Most Popular Models:

  1. SorcererLM 8x22b bf16 - 16000 Tokens Context
  2. Magnum-72b-v4 - 32000 Tokens Context
  3. 72B Qwen2.5 Kunou v1 - 32000 Tokens Context
Pros Cons:
Offers models that are unavailable on any other platform Recent major scandals involving the team's attempts to conceal changes to models, prioritizing speed over intelligence while misleading their user base
Up to 70B Models with 32K context for the lowest subscription cost (vs. competitors) and reliable uptime Models hosted on Infermatic often underperform compared to the same models on other platforms
Models oriented towards RP, detailed stories, character progression, etc Very small selection of models; new models are added slowly, and existing models may be removed without notice

Featherless: Recursal's Answer to OpenRouter

Featherless, the older sister of Wyvernchat is Recursal.ai's flagship product. Preferring to cast a wide net rather than get pigeonholed, they offer the largest selection of all uncensored models (even rivalling OpenRouter) from 12B to 72B at an output between 10-40 tokens per second, all for one flat fee. While it's Premium subscription is the most expensive on the market, in return users get something no other host offers: Full, unlimited use of DeepSeek V3/R1 uncensored. However, Recursal's corporate nature has been a source of concern for burnt users that business priorities will overshadow community interests in a tale as old as AI Dungeon.

Website: https://featherless.ai/
Discord: https://discord.com/invite/7gybCMPjVA
My referal code: https://featherless.ai/register?referrer=cybp9oor
Cost Structure: Subscription (10/20USD)
Most Popular Models:

  1. meta-llama/Meta-Llama-3.1-8B-Instruct - 16384 Tokens Context
  2. deepseek-ai/DeepSeek-R1 (Uncensored) - 32768 Tokens Context
  3. mistralai/Mistral-7B-Instruct-v0.2 - 8192 Tokens Context
Pros Cons:
Extremely large selection of models, including numerous roleplay-focused fine-tunes. Offers uncensored access to models like DeepSeek V3/R1, EVA, Kuanou, and Wizard As a Recursial product, Featherless is subject to corporate decisions and profit priorities, which may include changes that conflict with community interests
Provides popular presets for each model, saving users time in configuration While more stable than ArliAI, Featherless servers have been impacted recently by the intensive burdens that come with running major models such as DeepSeek as well as ongoing server upgrades.
Gives you the most popular presets for each model, which saves you some time searching Pricing is on the higher end compared to other subscription-based services, with lower average contexts than there competitors.
Featherless models have deep integration with Wyvernchat, giving them plug-and-play convenience While there's a community, the level of active engagement could be improved when compared too other hosts
Still new, falls victim to new team pitfalls Recently they've pulled thier heads out thier ass and server investments are paying dividends to reduce the breakdowns.

Addendum:

a. Its important to remember each user experiences will be different. What you want may not be what I was looking at.
b. If you found this write-up helpful and would like to continue learning, I welcome you to join us at AWR. Open to anyone who seeks to improve in their botmaking abilities: https://discord.gg/Ez53AXTVTy or https://discord.gg/awr
Thank you, Lufia, for providing a lot of the base points that I could use to jump off of to write up this primer. Check her out here: https://chub.ai/users/Lufia_ In addition to AWR, you can also find me at:

Edit Report
Pub: 17 Feb 2025 20:42 UTC
Edit: 30 Mar 2025 18:59 UTC
Views: 665