OpenAI makes GPT-4o Mini Fine-tuning Free, Mistral Releases Mistral Large 2
Created on July 25|Last edited on July 26
Comment
Starting today, organizations can fine-tune GPT-4o Mini for free until September 23, 2024. During this period, each organization is allotted 2 million tokens per 24 hours for training. Any tokens used beyond this limit will be charged at $3.00 per million tokens. Importantly, there are no charges for tokens used during training validation. For example, training a file with 100,000 tokens over 3 epochs would cost approximately $0.90 USD with GPT-4o Mini after the free period ends.
The Catch
GPT-4o Mini fine-tuning is available to developers in OpenAI’s Tier 4 and 5 usage tiers, which are the highest-priced tiers among OpenAI’s plans. OpenAI plans to gradually expand access to free fine-tuning to all tiers. Free fine-tuning will be offered now through September 23. Tier 4 requires a $250 payment and 14+ days since the first successful payment. Tier 5 requires a $1,000 payment and 30+ days since the first successful payment.
This is another interesting chess move by Sam Altman. By offering free fine-tuning of GPT-4o Mini, OpenAI is significantly lowering the barrier to entry for organizations to experiment with advanced AI models. This will likely increase buy-in from a diverse range of users, including startups, small businesses, and larger enterprises, who can now access and leverage this technology without initial financial constraints.
Mistral Large 2
Mistral AI has announced the release of Mistral Large 2, the next generation of its flagship model, offering significant advancements in code generation, mathematics, reasoning, and multilingual capabilities.
Improvements
Mistral Large 2 demonstrates substantial improvements in code generation, supporting over 80 programming languages, and shows marked enhancements in mathematical reasoning and multilingual support. It covers languages such as French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean. Designed for single-node inference, Mistral Large 2 features 123 billion parameters, maintaining high throughput. It delivers superior performance on evaluation metrics, achieving an 84.0% accuracy on the MMLU benchmark and offering cost-efficient serving capabilities.
Sharp as a Tack
The model's training focused on reducing "hallucinations" and enhancing the accuracy of responses. It is now better at acknowledging when it lacks sufficient information, reflected in improved performance on mathematical and problem-solving benchmarks. Mistral Large 2 also excels in instruction-following and alignment, showing improved abilities in handling precise instructions and long multi-turn conversations. Its multilingual proficiency ensures robust performance in business applications requiring diverse language support.
Available today via la Plateforme, under the name mistral-large-2407, users can test the new model on le Chat. Mistral Large 2's weights are hosted on HuggingFace, and it is available under the version 24.07. Mistral AI has partnered with leading cloud service providers like Google Cloud Platform, Azure AI Studio, Amazon Bedrock, and IBM watsonx.ai to make Mistral Large 2 widely accessible.
Sources:
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.