Which Open Source LLMs Are Your Go-To Choices?
06/06/2024 10:12:23When it comes to selecting the right language model for your frequent tasks, there's nothing like diving into a virtual discussion with AI enthusiasts. Here's a glimpse into which models are favored by professionals for different needs. This is just a friendly chat, not an official guide, but it's still interesting!
Most Popular Models
Llama Family Models: The Llama series remains a favorite. The Llama-3-70B is highly popular for everyday use, including coding. Llama-3-70B-Instruct, Llama-3-8B-Instruct, and Llama-3-8B are also preferable regularly.
Other Notable Mentions
- C4ai Command R Plus: Considered one of the best (but can be pricey). Highly appreciated for its cleverness, multilingual abilities, and being uncensored. Ideal for chatting in languages other than English.
- WizardLM-2 8x22B: Very smart and a good alternative in some situations to Llama 3 70B (🤔).
- Mixtral-8x7B-instruct and Phi-3-medium-128k-instruct: Good for general tasks.
For Coding
- CodeLlama - 70B: A favorite for complex coding tasks.
- Deepseek-Coder 6.7B Instruct: Excellent for code generation.
- Codestral-22B-v0.1: Some people find it the new champion in coding models.
RAG (Retrieval-Augmented Generation)
- Phi3:14b-medium-128k-instruct-f16: Fast for small to medium document collections, though it requires some tweaking.
- C4ai Command R Plus: Certainly! Dependable for RAG with small to medium collections.
RP (Role Play) Purposes
- Midnight-Miqu and Fimbulvetr-11B-v2: Reliable and time-tested choices for role-playing purposes, trusted by many users for their consistency and effectiveness.
Everyone has their go-to models, and the choice often depends on specific needs. What are your go-to choices? Share your favorites!
Recent Blog Posts
-
2024-08-03
Original data from HuggingFace, OpenCompass and various public git repos.
Release v2024072803