Staying current in the rapidly evolving landscape of artificial intelligence can feel overwhelming, especially with new AI models launching almost daily. Here’s a roundup of six significant models that have emerged recently, offering cutting-edge advancements and features to watch.
### Apple’s OpenELM
**Is it open source?** Yes
Apple has made its entry into the open-source AI space with OpenELM, marking a significant step for the tech giant known for its proprietary software. OpenELM, which stands for Open-source Efficient Language Models, includes models ranging from 270 million to 3 billion parameters.
Apple's OpenELM comes in two variations: a pre-trained version and an instruction-tuned variant designed to respond effectively to natural language commands. Each model leverages CoreNet, Apple’s state-of-the-art deep neural network library, trained on an extensive dataset comprising 1.8 trillion tokens from publicly available sources like RefinedWeb.
The architecture of OpenELM incorporates innovative layer-wise scaling techniques that enhance response accuracy by optimizing parameter allocation within layers.
#### Accessing OpenELM
Both versions of OpenELM can be found on Hugging Face, while the CoreNet library is accessible on GitHub. While the models can be integrated into commercial applications, users must adhere to stringent licensing terms, including avoiding any implication of endorsement by Apple.
### Snowflake’s Arctic
**Is it open source?** Yes
Snowflake has recently launched Arctic, an enterprise-centric large language model boasting 17 billion parameters. Designed for efficiency, Arctic is capable of executing tasks such as code generation while maintaining a low operational cost compared to other models.
Arctic allows for the creation of custom models tailored to specific enterprise needs, performing on par or better than other models, including Meta’s Llama 3. This model aligns with Snowflake’s principles of being truly open, providing users with access to its weights, code, and the datasets used for training.
#### Accessing Snowflake Arctic
Arctic is available for download from Hugging Face and can also be accessed through various cloud providers, including AWS, Azure, and Nvidia's API catalog.
### Microsoft’s Phi-3 Mini
**Is it open source?** Yes
Microsoft’s Phi-3 Mini is a compact model with 3.8 billion parameters that surpasses larger models in reasoning, coding, and math capabilities. Notably, this model can handle a context window of up to 128K tokens, making it adaptable for edge applications like smartphones and industrial sensors.
Phi-3 Mini is instruction-tuned, enabling it to respond to user commands straight away, which streamlines deployment across multiple platforms.
#### Accessing Phi-3 Mini
Users can access Phi-3 Mini through Azure AI Studio, Hugging Face, Ollama, and Nvidia’s Nim platform, which facilitates easy deployment options.
### Meta’s Megalodon
**Is it open source?** Yes
Meta's Megalodon, named after the ancient shark species, aims to tackle large-scale and complex tasks with a user-friendly interface that manages lengthy inputs and maintains context. This model is engineered for speed and scalability, intended to enhance performance in data-intensive applications.
While specific parameter counts have not been disclosed, Megalodon is designed to operate efficiently, fast-tracking the processing of extensive datasets.
#### Accessing Meta Megalodon
Users can find Megalodon on GitHub, where a Discord channel is also available for troubleshooting and community support.
### Mistral AI’s Mixtral 8x22B
**Is it open source?** Yes
Released by French startup Mistral AI, Mixtral 8x22B stands out with 141 billion parameters yet only a 218GB file size, making it manageable for consumer-grade systems. This model uses a mixture of expert (MoE) architecture, achieving high performance while ensuring cost-efficiency.
Mixtral 8x22B boasts capabilities similar to Meta’s Llama 2 and OpenAI’s GPT-3.5, offering users flexibility to use it in various proprietary applications due to its Apache 2.0 license.
#### Accessing Mixtral 8x22B
Interested users can demo Mixtral through Together AI’s API or explore its capabilities on Perplexity.ai. The model can be downloaded via a specific link shared through Mistral’s social channels.
### OpenAI’s GPT-4 Turbo
**Is it open source?** No
Rounding out this list is GPT-4 Turbo, the latest advancement from OpenAI. This powerful model is now available to premium ChatGPT subscribers, offering significant improvements in coding, math, and reasoning at a reduced operational cost.
Equipped with an impressive 128K token context length, GPT-4 Turbo can handle extensive information inputs and outputs, making it an invaluable tool for complex project automation and visual coding.
#### Accessing GPT-4 Turbo
Users can experience the capabilities of GPT-4 Turbo across OpenAI's platform, integrated with dynamic features that facilitate improved interactions and outputs.
---
As AI continues to advance, these models represent significant footholds in an increasingly competitive landscape, catering to diverse needs from enterprise solutions to everyday applications.
Ranking
