Simon Willison’s Weblog: Un Ministral, des Ministraux

Source URL: https://simonwillison.net/2024/Oct/16/un-ministral-des-ministraux/
Source: Simon Willison’s Weblog
Title: Un Ministral, des Ministraux

Feedly Summary: Un Ministral, des Ministraux
Two new models from Mistral: Ministral 3B and Ministral 8B (joining Mixtral, Pixtral, Codestral and Mathstral as weird naming variants on the Mistral theme.

These models set a new frontier in knowledge, commonsense, reasoning, function-calling, and efficiency in the sub-10B category, and can be used or tuned to a variety of uses, from orchestrating agentic workflows to creating specialist task workers. Both models support up to 128k context length (currently 32k on vLLM) and Ministral 8B has a special interleaved sliding-window attention pattern for faster and memory-efficient inference.

Mistral’s own benchmarks look impressive, but it’s hard to get excited about small on-device models with a non-commercial Mistral Research License and a contact-us-for-pricing Mistral Commercial License, given the existence of the extremely high quality Llama 3.1 and 3.2 series of models.
These new models are also available through Mistral’s la Plateforme API, priced at $0.1/million tokens (input and output) for the 8B and $0.04/million tokens for the 3B.
Via Hacker News
Tags: mistral, llms, ai, generative-ai

AI Summary and Description: Yes

Summary: The text discusses the introduction of two new AI models from Mistral—Ministral 3B and Ministral 8B—highlighting their advancements in knowledge, commonsense reasoning, and efficiency, particularly with regards to large language models (LLMs). This information is highly relevant to AI professionals looking for cutting-edge developments in LLM technology.

Detailed Description: The new AI models introduced by Mistral represent significant advancements in the field of large language models (LLMs):

– **Models Overview**: Ministral 3B and Ministral 8B are the latest additions in Mistral’s series of models, which include Mixtral, Pixtral, Codestral, and Mathstral. These models explore a new frontier in several key areas:
– **Knowledge and Commonsense Reasoning**: These models have been designed to enhance their understanding and commonsense reasoning capabilities, making them more efficient for various applications.
– **Efficiency**: Both models aim to provide greater efficiency particularly within the sub-10 billion parameter category, allowing for effective use in constrained environments.

– **Technical Specifications**:
– **Context Length**: They support an impressive context length of up to 128k, which is currently a significant improvement over the 32k supported by vLLM.
– **Inference Methodology**: The Ministral 8B model features a unique interleaved sliding-window attention pattern, contributing to faster and more memory-efficient inference.

– **Benchmarks and Usage**: Despite Mistral’s optimistic benchmarks, there are concerns about the practical application of smaller models in light of competition from higher-quality models like Llama 3.1 and 3.2. The licensing structures (non-commercial and commercial) also add a layer of complexity for potential users.

– **Access and Pricing**: Both models can be accessed via Mistral’s la Plateforme API, with pricing set at:
– **Ministral 8B**: $0.10/million tokens for both input and output.
– **Ministral 3B**: $0.04/million tokens, providing various pricing options depending on usage.

The release of these models signals continued innovation in the LLM space, with professionals in AI and cloud computing benefitting from insights into model performance and applications, particularly for orchestrating workflows or creating specialized task workers.