Model Catalog

14B version of the code-specific Qwen 2.5 for code generation, code reasoning and code fixing.

32B version of the code-specific Qwen 2.5 for code generation, code reasoning and code fixing.

3B version of the code-specific Qwen 2.5 for code generation, code reasoning and code fixing.

A tiny and speedy Llama model from Meta, optimized for multilingual dialogue use cases.

The new and small Llama model from Meta, optimized for multilingual dialogue use cases, including agentic retrieval and summarization tasks.

Code-specific LLMs for code generation, code reasoning and code fixing, supporting context length of up to 128K.

An LLM specializing in instruction following, generating long texts (over 8K tokens), understanding structured data (e.g, tables), and generating JSON structured outputs.

Yi Coder is a Llama fine-tune with expanded size, trained for code. Supports max 128K tokens and "52 major programming languages".

A fine-tune of Meta's Llama 3.1, Hermes is further trained on hand-curated datasets as well and synthetic data. Excels in dialogue and code generation.

InternLM 2.5 offers strong reasoning across the board as well as tool use for developers, while sitting at the sweet spot of size for those with 24GB GPUs.

The original LLaVA vision-enabled model, supporting image input and textual instruction following.

The latest in Meta's long-running Llama series, Llama 3.1 is another jack of all trades and master of some, now in 8 languages and up to 128k tokens.

A slightly larger 12B parameter model from Mistral AI, NeMo offers a long 128k token context length, advanced world knowledge, and function calling for developers.

"Google's Llama", Gemma benefits from Google's experience training its flagship Gemini model to provide excellent performance on low power or for autocompletion/drafting tasks.

A scientific specialist finetune of Mistral AI's popular 7B model, Mathstral excels at STEM chats and tasks.

The mid-sized option of the Gemma 2 model family. Built by Google, using from the same research and technology used to create the Gemini models

A HuggingFace original model, SmolLM lives up to its name in size and will fit on just about any device. Slightly larger option at 1.7B parameters is also available.

The large option of the Gemma 2 model family. Built by Google, using from the same research and technology used to create the Gemini models

Mistral AI's latest coding model, Codestral can handle both instructions and code completions with ease in over 80 programming languages.

Microsoft's latest Phi Mini model supports a whopping context length of 128k tokens in a small size, offering extremely long chats for cheap.

The younger sibling of the GPT-4-beating 236B Deepseek Coder V2 model, this model also comes out strong with support for 338 different languages!

A small model from Alibaba's Qwen2 family that punches above its weight in mathematical and multi-step logical reasoning.

Promising 27 languages and lightning fast responses, this is the smallest entry in Alibaba's Qwen2 family that scales up to 72B parameters.

Aya 23 is an open weights research release of an instruction fine-tuned model with highly advanced multilingual capabilities.

One of the most popular open-source LLMs, Mistral's 7B Instruct model's balance of speed, size, and performance makes it a great general-purpose daily driver.

From the team behind Stable Diffusion, this small code model offers an excellent coding assistant for those with lighter hardware.

Able to chat in more than 10 languages, Cohere AI's Command-R is optimized for RAG but can perform well in any task.

Also coming in 3B, 15B, and Chat versions, the StarCoder2 family offers a diverse portfolio of local coding assistants.

Promising comparable performance to GPT-4 on mathematical reasoning, DeepSeek Math also offers the ability to write code to solve and prove mathematical problems.