Intelligence

Large Language Models

ChatGPT (OpenAI)
Gemini (Google) [formerly known as Bard]

Inference

llama.cpp
ollama
LM Studio (`lmstudio-js` (TypeScript SDK))
Jan
vllm

Models

I prefer the non 'instruct' variants.. I'm really not sure which are the best,
but all of these are my favorite ones (older ones have been and will be removed, from time to time).
To use these ones with the transformers above, look at the Tools section.

BUT I'd *maybe* recommend you the abliterated ones; even though they're not *that* up2date.. you decide.
PS: You can look to the source of this file for some more models.. I sort(ed) them out, leaving only the newest and biggest ones visible here.

DeepSeek-R1
DeepSeek-V3-Base
Mistral-Large-Instruct-2407
Mixtral-8x22B-v0.1
Hermes 3
Llama-3.3-70B-Instruct
Qwen2.5-72B
MiniMax-Text-01

Abliterated Models


It's about uncensoring LLMs ... the "easy" way (in a "global" form, not as usual by changing the prompt(s)). Find out more here:

Uncensor any LLM with abliteration #1
Uncensor any LLM with abliteration #2
Refusal in LLMs is mediated by a single direction #1
Refusal in LLMs is mediated by a single direction #2
Demo of bypassing refusal

Tools

Created by myself, to make it easier for me to handle all the models.

convert-hf-to-gguf.sh (.sh)
hfdownloader.sh (.sh)
hfget.sh (.sh)

Links

Large Language Model Course
Neural Network Zoo (Image)
"What are the risks from Artificial Intelligence?"
"Machine learning in a few clicks"
AutoDAN-Turbo: A Lifelong Agent for Strategy Self-Exploration to Jailbreak LLMsGitHub
Genesis
LLM Visualization
Spreadsheets are all you need
NodeJS library forLlama Stack
The GPT-3 Architecture, on a Napkin
Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs
Microsoft KBLaM: Knowledge Base augmented Language Model
MICrONS Explorer: A virtual observatory of the cortex

`Norbert`

This is my very own artificial intelligence.

It's totally based on pure byte code, nothing with tokens or smth. like it (like current LLMs are based on).
And such input is processed abstract. The last output is then also injected once again as second, parallel input byte (some feedback loop).

And here are some screenshots of the process itself and two helper utilities I've created especially for this reason; also my dump.js.

Example screenshot
Example `learn` screenshot
Debugging the intelligent output
Bit testing ..