Mistral: Our first AI endpoints are available in early access (mistral.ai)

Article summary

Mistral AI has announced the early access release of its first AI endpoints, including three chat endpoints and an embedding endpoint. The endpoints have different performance and price tradeoffs, with the highest-quality endpoint serving a prototype model that outperforms ChatGPT 3.5. The models are open-sourced and can be downloaded, and the API follows the specifications of a popular chat interface. The company is progressively ramping up its capacity and offers a Python and Javascript client library to query its endpoints.

Main themes

• AI endpoints
• open-source models
• pricing and value proposition
• reliability and portability
• model quality and comparison
• vendor lock-in and risk mitigation
• local deployment and power consumption
• compression and quantization of models
• insurance against model changes

What commenters say

• The pricing of Mistral's API suggests that the company is confident in the quality of its models, particularly the Mistral-medium model.
• Some commenters believe that the value proposition of paying for an open-source model is not clear, as it may be possible to host the model oneself or pay a third-party to do so.
• Others argue that the reliability and portability of an open-source model make it a more attractive option, even if it is not the most advanced model available.
• There is debate over whether the Mistral-medium model is substantially better than GPT-3.5, with some commenters suggesting that the pricing may be due to factors other than model quality.
• The open-sourcing of the models mitigates the risk of vendor lock-in and allows users to take the model and run it themselves if needed.
• Some commenters are skeptical of the idea that a large language model can be run locally on a phone, citing power consumption and latency concerns.
• The potential for compressing and quantizing the models to reduce their size and increase their speed is seen as an exciting development with potential applications.
• The use of open-source models can provide insurance against model changes and allow users to continue using a model even if the host decides to stop offering it.