10:30Machine-translated from Azerbaijani by AIRead in Azerbaijani
Mistral Unveils New AI Model With 1 Trillion Parameters
Mistral unveiled its Mistral Large 4 artificial intelligence model with more than 1 trillion parameters this week. Parameters are internal values learned by a model, and their number indicates the model’s size. The company positions it as a cutting-edge solution among open-weight models. The model is still in open testing for research purposes and runs on Mistral’s API service. The files needed to run the model independently are scheduled to be released by the end of October.
In testing by Artificial Analysis, Mistral Large 4 scored 38 points on the Intelligence Index. This is the highest score among models developed outside the US and China. However, at least 5 open Chinese models performed better. MiMo-V2.6-Pro scored 46 points, while GLM-5.3 scored 45. Kimi K3 received 44 points, and GLM-5.3-Flash scored 42. DeepSeek V4.1 Flash recorded 39 points.
Mistral Large 4 uses a mixture-of-experts architecture that activates only the components required for a given task. The model has 1.05 trillion parameters in total, but only 49–52 billion of them are active at the same time. Its context window is 512,000 tokens, meaning the model can process very long texts in a single request. It accepts both text and images.
The API has separate pricing for 1 million input tokens. This volume of input costs $1.36 (approximately AZN 2). The same number of output tokens costs $4.18 (approximately AZN 7). Prices are discounted by 50% for the first 2 weeks.
The model scored 50 points on the overall cybersecurity index, matching GLM-5.3-Flash. It outperformed all competitors in the CyberGym-E2E-AA test with a score of 81.7%. However, its results were significantly lower in other security tests. It scored 16% in the DeepsecBench-AA test and 51% in the CWE-Bench-AA test.
Completing 1 task costs $1.13 (approximately AZN 2) under standard pricing. The discount remains in effect for the first 2 weeks. During that period, the price is $0.57 (approximately AZN 1). This is more than 4 times as expensive as open models delivering similar results. Mistral Large 4 generates 116.1 tokens per second, and the first part of its response appears in 1.46 seconds. The API accepts 100 images in a single request. Previous Mistral models had a limit of 8 images.
Mistral Large 4 was developed in about 2 months at data centers in Europe using 3,800 Nvidia Grace Blackwell graphics processors. The company plans to use the model as a foundation for developing specialized solutions in the future. Mistral previously raised €3 billion (approximately AZN 5.73 billion) in Series D funding. Once the model’s files are released, it can be run in private cloud systems and on companies’ own servers.