Artificial Intelligence

Mistral Large 4 Trails Some Chinese Open Models in Tests

Mistral has launched its trillion-parameter Large 4 model in Research Public Preview, securing strong performance in cybersecurity evaluations while trailing several Chinese open models on independent intelligence indices.

Radwan AbdulhadiEditor and technology writer

Published: October 10, 2026 · Updated: October 10, 2026 · 3 min read

Mistral Large 4 Trails Some Chinese Open Models in Tests — Image: Tom's Hardware

Launch and Independent Rankings

French artificial intelligence company Mistral has released its trillion-parameter Large 4 model in Research Public Preview, with plans to release the weights of the 1T parameter model—featuring 49B active parameters—by the end of October. According to independent AI benchmarking firm Artificial Analysis (AA), the mixture-of-experts model scored 38 on the Artificial Analysis Intelligence Index v4.3.2. This performance positions Mistral Large 4 as the most intelligent model available from outside the United States and China.

Despite earning recognition as the leading model from Europe, independent tests show that Mistral Large 4 trails several Chinese open models on the broader intelligence index. At least five Chinese open models scored higher, including DeepSeek V4.1 Flash (max) at 39, GLM-5.3-Flash at 42, Kimi K3 (max) at 44, GLM-5.3 (max) at 45, and MiMo-V2.6-Pro at 46. While MiMo-V2.6-Pro is comparable in total size to Mistral Large 4, GLM-5.3-Flash is significantly smaller at approximately 320B total parameters.

Cybersecurity and Specialized Performance

While trailing on general intelligence indices, Mistral Large 4 achieved notable results in specialized security evaluations. The model scored 50 on the Artificial Analysis Cyber Index, placing it level with GLM-5.3-Flash and behind MiMo-V2.6-Pro. Its strongest result emerged on the CyberGym-E2E-AA end-to-end cybersecurity measure, where it scored an 81.7% to 82% rating, outperforming several competitors and surpassing OpenAI's closed GPT-6 Luna.

Mistral stated that the model is state-of-the-art among open models on enterprise workloads such as cybersecurity, finance, and law. CEO Arthur Mensch noted in Abu Dhabi that the model sits above Chinese models on certain aspects, including cyber capabilities. Artificial Analysis projects that once the model's weights officially ship at the end of October, it will rank among the top three open-weights models on the Cyber Index.

Mistral AI on a desktop — (Image credit: Getty Images / Bloomberg) · Source: Tom's Hardware

Cost, Pricing, and Efficiency

In terms of operational economics, Mistral Large 4 comes in at a higher cost compared to similar-intelligence open-weights alternatives. At standard list pricing, it costs $1.13 per Artificial Analysis task, driven by standard pricing of $1.36 and $4.18 per 1 million input and output tokens respectively, alongside $0.14 per 1 million cached input tokens. For its first two weeks, Mistral is offering a 50% launch discount that reduces its cost per task to $0.57.

Despite the discount, the model remains more costly than comparable Chinese open offerings like GLM-5.3-Flash at $0.25 per task and DeepSeek V4.1 Flash at $0.27. Artificial Analysis reports that the elevated cost partly stems from requiring 200 million output tokens to run its index evaluations, compared to a median of 81 million tokens for other models.

Hardware Infrastructure and Training

Mistral Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs housed within Mistral's own datacenters in Europe over roughly a two-month period. This infrastructure marks a hardware upgrade from the 3,000 Nvidia H200 accelerators used to train Mistral Large 3, which was released previously with 675B total parameters and 41B active parameters.

In addition to its hardware foundation, the model features a context window of 512k tokens, supports text and image inputs with text output, and demonstrates strong document and image reasoning. Its visual capabilities are driven in part by a 1.6B-parameter vision encoder, allowing the API to accept 100 images per request—a substantial increase from the 8-image limit found in previous Mistral architectures.