One point behind, half the tokens: Mistral Large 4 ranks second on the open-source coding leaderboard

On October 31, Mistral released the weights for Mistral-Large-4-1T-A52B on HuggingFace. The chat entry point is also open, with the address at the end of this article. A coding agent is an AI tool that can read repositories, run terminals, and modify code. Two numbers in this announcement are more worth noting: a Coding Agent Index score of 48 and an average total of 8.5M tokens per task.

First, the ranking. Artificial Analysis's Coding Agent Index v2.0 combines FrontierCode v1.1, Terminal-Bench 4.0, and SWE-Atlas-QnA-AA into a single score. Mistral Large 4 scores 48 in Mistral Vibe, ranking second among open-weight models. Ahead is GLM-5.3 in OpenCode with 49; behind is DeepSeek V4.1 Flash 1017 in DeepSeek Harness with 46. Claude Code remains first overall with 62. Mistral's result is sixth overall and second among open-source models.

The gap is only 1 point, but token usage differs by a lot. According to Artificial Analysis, Mistral Large 4 uses about 8.5M total tokens per task, GLM-5.3 about 15.9M, and DeepSeek V4.1 Flash about 16.7M. By comparison, Mistral is roughly half that of its rivals. When a coding agent performs long tasks, tokens are both the bill and the latency. When scores are close, this difference is more interesting than being 1 point ahead.

But don't treat the index as everything. The Coding Agent Index measures coding-agent tasks, and scores are affected by the harness, prompts, and toolchain. Mistral Large 4 runs in Mistral Vibe, GLM-5.3 runs in OpenCode, and DeepSeek runs in its own harness. The leaderboard tries to control variables, but real projects have more variables. The nine-benchmark chart also shows that different models trade places across different dimensions. When choosing a model, first consider which benchmark your task resembles, then look at the token bill.

The 1T-A52B in the name is also worth mentioning. Under common naming conventions, this is an MoE model with 1T total parameters and 52B activated parameters; refer to the model card for specifics. Total parameters are large while activated parameters are relatively small, meaning not every parameter does work during inference. But that doesn't mean it can fit on a small machine. The deployment barrier of 1T total parameters is still there, and quantization, VRAM, and throughput all need to be calculated. For ordinary users, the easiest option is to try it directly at chat.mistral.ai. For developers, the HuggingFace weights page is better for checking the model card, license, and inference requirements.

Someone joked in a reply about a 'fat model smell.' The complaint isn't unfair. The larger the model, the heavier the engineering burden. The counterintuitive point Mistral offers this time is that large models aren't necessarily more expensive in tokens. With a similar score and fewer tokens, it has pulled competition on the open-source coding leaderboard into the efficiency column.

Red Router's comment is direct: GLM-5.3 still holds the open-source top spot by a 1-point margin. The gap is very small and could change in the next evaluation. If you really need to choose, don't just look at the ranking. Run a round with your repository, your terminal tasks, and your budget. The weights are here: https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B . The trial entry point is here: https://chat.mistral.ai/chat . Leave Mistral a piece of feedback too. Leaderboard details are on Artificial Analysis: https://artificialanalysis.ai/models/mistral-large-4 .