Mistral open source 8X22B AI expert hybrid model

Mistral released its latest artificial intelligence (AI) model, the 8X22B, on Wednesday. The expert hybrid AI model follows the 8X7B model released in December 2023 and provides a larger number of parameters. The company, known for its fully open source artificial intelligence models, unconventionally released its latest model without any accompanying announcement or blog post. While the AI ​​company itself doesn’t provide benchmarks, Hugging Face community users tested it and posted benchmark scores for the 8X22B model. The results appear to close the gap with OpenAI and Google’s closed models.

Mistral’s official X (formerly Twitter) account freed 8X22B AI model continues its unconventional way of releasing AI models without announcement via torrent magnet links, directly for people to download. Mistral is also one of the only truly open source platforms that makes not just the weights open, but the entire architecture. However, it is important to note that most devices do not have the ability to run powerful AI models on the device, which may cause the device to malfunction. The total file size is 262GB.

Mistral’s 8X22B is an auto-complete AI model. These are typically different from the AI ​​model’s instructions or chat variations. OpenAI’s ChatGPT and Google’s Gemini AI are chat models, while Meta’s Code Llama 7B and 13B are command models. The chat model is able to understand natural language and contextual queries to provide the correct response. Instruction models are mainly used by developers who ask AI models to perform specific tasks. Instead, the autocomplete model completes the sentences provided in the prompt.

Early benchmark scores based on hugging faces community posts Showing that the latest Mistral model offers significant upgrades over its predecessor and narrows the gap with other major models. On the Hellaswag benchmark, the Mistral 8X22B scored 88.9, just behind GPT-4 (95.3), Claude 3 Opus (95.4) and Gemini 1.5 Pro (92.5). However, its score exceeded GPT-3.5 (85.5) and Gemini 1.0 Ultra (87.8).

Comparing to the benchmark of Mistral’s own models, the 8X22B falls between the Mistral medium and Mistral large models. It might even be similar to Mistral-large as commands or chat variations are released later.