Mistral Large 4 vs Qwen3.8 Max
Mistral Large 4 (ML4, "Le Chonk"), announced October 6, 2026, is Mistral AI's open-weight flagship at 1.05T total parameters. Qwen3.8 Max is the 2.4T-total-parameter model from the Chinese side of the field. The two are the heavyweights of the open race, but reported benchmark margins between ML4, DeepSeek V4 Pro, and Qwen3.8 Max are narrow: no single model holds a decisive lead across the board.
Head-to-head specs
| Spec | Mistral Large 4 | Qwen3.8 Max |
|---|---|---|
| Total parameters | 1.05T | 2.4T |
| Active parameters per token | 49B (MoE) | Not disclosed in our sources |
| Architecture | MoE | Not disclosed in our sources |
| Context window | 1M tokens | Not disclosed in our sources |
| Preview API pricing | $1.36 / 1M input tokens, $4.18 / 1M output tokens | Not disclosed in our sources |
| Open weights | Expected October 27, 2026 (Reuters reported); custom Mistral license expected | Not disclosed in our sources |
| Training infrastructure | Trained from scratch on ~3,800 Nvidia Grace Blackwell GPUs in Mistral's own European datacenters | Not disclosed in our sources |
| Reported cyber benchmarks | 93% Cybench, 82% CyberGym-E2E (vendor-reported) | Not disclosed in our sources |
Where Mistral Large 4 stands out
- Efficiency per parameter. At 1.05T total parameters, ML4 is less than half the size of Qwen3.8 Max's 2.4T total. Fewer total parameters generally means cheaper serving and hosting costs for a given throughput, which matters for anyone deploying the weights themselves once the open release lands.
- Cyber benchmark focus. Mistral's vendor-reported numbers for ML4 are 93% on Cybench and 82% on CyberGym-E2E. CEO Arthur Mensch said the model is above Chinese models on certain aspects, including cyber.
- European data sovereignty. ML4 was trained from scratch in Mistral's own European datacenters, on ~3,800 Nvidia Grace Blackwell GPUs. For organizations with EU data-residency or sovereignty requirements, that provenance is a concrete differentiator.
- Language coverage. ML4 supports 160+ languages, including every official EU language, plus native multimodal input through a 1.6B vision encoder.
- External validation. Artificial Analysis places the ML4 preview between DeepSeek V4.1 Flash and OpenAI's GPT-6 Luna on its intelligence leaderboard, and Mistral claims ML4 is the best open-weight model from the US or Europe on aggregated benchmarks and competitive with the strongest open models worldwide.
Where Qwen3.8 Max may have an edge
Qwen3.8 Max's 2.4T total parameter count is the largest of the three models in this comparison set (ML4 at 1.05T, DeepSeek V4 Pro at 1.65T). Scale alone does not decide capability — the reported margins across the three are narrow — but it is the biggest open-class model by total parameter count. Beyond that, our sources do not disclose its active parameter count, pricing, or benchmark figures, so any deeper claim would be speculation.
Verdict
The margins are narrow and no single model holds a decisive lead across the board, so this is not a one-sided race. ML4 offers the smaller, cheaper-to-serve footprint (less than half Qwen3.8 Max's total parameters), a strong cyber benchmark story backed by Mensch's claim of leading Chinese models on cyber, EU-trained provenance for sovereignty-sensitive deployments, and broad multilingual coverage. Qwen3.8 Max counters with the largest total parameter count of the three.
The decision depends on your workload and deployment constraints: cost-efficient hosting and EU data residency point toward Mistral Large 4; raw scale points toward Qwen3.8 Max. If pricing and benchmarks for Qwen3.8 Max are published, revisit this comparison — the current gap in disclosed figures is the biggest unknown.