/aicg/ meta

Comparison between the different services/models and frontends used by /aicg/.
These ratings aren't gospel. They're the opinion of one anon who tried to incorporate suggestions and push no agenda.

Services/models

If you're curious about a service, the OP should have the information you need.

🏆 - Best
🥈 - Good
🥉 - Usable
🌱 - Bad
⛔ - Worst
🔀 - Depends
NSFW - general ERP-readiness (taking into account the need to jailbreak)
SFW - general RP-readiness
MEMORY - more size than retention, more memory = higher price for both cloud and local
CREATIVITY - low only for dumb assistants, can be increased with a CoT prompt
COMPLEX INSTRUCTIONS - stats tracking, logical conditions, multiple characters, etc.
PRICE - per token, per month, or for running local models one time purchase or per hour rental of the hardware
EASE OF USE - OAI keys or all proxies are easy, but still require JBs; Horde and subscription based local/NAI offer inbuilt nobrain presets; running local is not that hard because there's enough guides; anything can require shuffling prompt parts and finding the best temperature and other settings
/lmg/ - local models in general, the ones you can download and run offline, the best ones in the rankings

Notes:

• Gemini API: free access is rate limited.
• Cohere API: free access is rate limited.
• AI21 Jamba API: free trial is limited.
• NovelAI models:
  • Not intended for chat usage and not recommended.
  • But since they're bundled with Image Gen sub, you might want to experiment.
• Local models:
  • Check out /lmg/
  • Small brain/short memory models can be run on very basic hardware, but size+memory+speed vs. requirements curve is very steep.
  • Settings and presets are shared as a single file for easy import and setup.
• OpenAI o1/o3: thinking models by OpenAI aren't worth using for RP
• DeepSeek-V3: usable but at the same time has severe issues.
• DeepSeek-R1: enjoyable but at the same time has annoying issues.

Quick notes on how to access each:

• GPT:
  • Official API, revendors.
  • Or use a reverse proxy that has it. (agnai video guides / tavern guide)
• Claude: Official API, revendors.
• Claude (AWS): Pay for it https://aws.amazon.com/bedrock/pricing/
• Gemini API: Official API through AI Studio (https://makersuite.google.com/app/apikey).
• Mistral: Official API (https://console.mistral.ai).
• Cohere: Official API (https://dashboard.cohere.com).
• Grok: Official API (https://console.x.ai/).
• DeepSeek: Official API (https://platform.deepseek.com).
• Jamba: Official API (https://studio.ai21.com).
• /lmg/
  • Use a reverse proxy or Horde.
  • Pay for a subscription that offers API access. Featherless (https://featherless.ai/#pricing) might be a good choice.
  • Some services offer free limited API access to small models.
  • Or learn to run it locally, on Colab or via GPU rental services.

Most models are available on OpenRouter (https://openrouter.ai/models) or from other revendors.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47

██ ███    ██ ███████  ██████  ██████  ███    ███  █████  ████████ ██  ██████  ███    ██ 
██ ████   ██ ██      ██    ██ ██   ██ ████  ████ ██   ██    ██    ██ ██    ██ ████   ██ 
██ ██ ██  ██ █████   ██    ██ ██████  ██ ████ ██ ███████    ██    ██ ██    ██ ██ ██  ██ 
██ ██  ██ ██ ██      ██    ██ ██   ██ ██  ██  ██ ██   ██    ██    ██ ██    ██ ██  ██ ██ 
██ ██   ████ ██       ██████  ██   ██ ██      ██ ██   ██    ██    ██  ██████  ██   ████ 


██████  ███████ ██       ██████  ██     ██                                              
██   ██ ██      ██      ██    ██ ██     ██                                              
██████  █████   ██      ██    ██ ██  █  ██                                              
██   ██ ██      ██      ██    ██ ██ ███ ██                                              
██████  ███████ ███████  ██████   ███ ███                                               


 ██████  ██    ██ ████████ ██████   █████  ████████ ███████ ██████                      
██    ██ ██    ██    ██    ██   ██ ██   ██    ██    ██      ██   ██                     
██    ██ ██    ██    ██    ██   ██ ███████    ██    █████   ██   ██                     
██    ██ ██    ██    ██    ██   ██ ██   ██    ██    ██      ██   ██                     
 ██████   ██████     ██    ██████  ██   ██    ██    ███████ ██████                      


██   ██ ███████ ██████  ████████     ███████  ██████  ██████                            
██  ██  ██      ██   ██    ██        ██      ██    ██ ██   ██                           
█████   █████   ██████     ██        █████   ██    ██ ██████                            
██  ██  ██      ██         ██        ██      ██    ██ ██   ██                           
██   ██ ███████ ██         ██        ██       ██████  ██   ██                           


██   ██ ██ ███████ ████████  ██████  ██████  ██  ██████  █████  ██                      
██   ██ ██ ██         ██    ██    ██ ██   ██ ██ ██      ██   ██ ██                      
███████ ██ ███████    ██    ██    ██ ██████  ██ ██      ███████ ██                      
██   ██ ██      ██    ██    ██    ██ ██   ██ ██ ██      ██   ██ ██                      
██   ██ ██ ███████    ██     ██████  ██   ██ ██  ██████ ██   ██ ███████                 


 ██████  ██████  ███    ██ ████████ ███████ ██   ██ ████████                            
██      ██    ██ ████   ██    ██    ██       ██ ██     ██                               
██      ██    ██ ██ ██  ██    ██    █████     ███      ██                               
██      ██    ██ ██  ██ ██    ██    ██       ██ ██     ██                               
 ██████  ██████  ██   ████    ██    ███████ ██   ██    ██                               

Frontend (service support) (outdated)

• Character Card V3 Implementation Tracking: https://rentry.org/zdwfrgqq

Three frontends are currently recommended, Silly, Agnai, and Risu. All have a lot of different features and are improving fast.

Miku is noteworthy, but difficult to compare. It uses sentiment classification to show different sprites for each message, and text-to-speech, supporting OAI, Kobold, Whisper, AzureTTS, ElevenLabs, and NovelAITTS.

Gemini works on all frontends if using it via reverse proxy. Mistral API supports OpenAI input (using the official API like a reverse proxy).

Frontend (features) (outdated)

• Only agnai has support for choosing which character speaks at any time in multi-user rooms, although on the flipside it doesn't support fully automatic mode
• (ext) means you have to install the Silly extras. They can be a little complicated to setup for non-technical users.

Outdated uninteresting information

• Poe: Create a bot over at https://poe.com/. Example: https://poe.com/coolreisen
• Slaude: it's still viable for sfw purposes for anons with actual enterprise accounts
• Bing: there are ways to use this version of GPT-4 for free, but it's "filtered"
• Clewd: there are ways to use claude.ai version of Claude for free, but it's "filtered"
• text-davinci-003 has been removed as OpenAI is discontinuing it.
• Davinci-3: Use an OpenAI key directly in SillyTavern/agnai. You can use your own or buy one (search "buy openai key" on yandex).
• Turbo: Same as davinci-3 but you can also use a reverse proxy.
• Gemini Advanced: the rating is provided for Gemini 1.0 Ultra, which was served to subscribers in the beginning. Pay for a subscription.
• Consensus on Gemini 1.5 Pro is: better than 3.5 Turbo, worse than 4 Turbo.
• Grok API: free credits until the end of 2024.