Mistral AI Releases Mistral Large 4 in Public Preview: 1.05T Parameter MoE with 1M Token Context

Launch Details and the 1.05-Trillion-Parameter Architecture

On October 6, 2026, French artificial intelligence laboratory Mistral AI officially launched Mistral Large 4 (ML4), internally codenamed 'Le Chonk,' into public preview. Developers can access the preview through the API on Mistral Studio immediately, with an open-weight release scheduled for the end of October 2026, reportedly targeted for October 27.

Architecturally, Mistral Large 4 is a natively multimodal sparse Mixture-of-Experts (MoE) system encompassing approximately 1.05 trillion total parameters. Despite its massive total footprint, the model routes execution sparsely, activating 49 billion parameters per token during inference, supported by a 1.6-billion-parameter vision encoder. The model natively processes a context window of up to 1 million tokens.

Mistral AI confirmed that ML4 was pretrained from scratch across more than 160 languages using a cluster of 3,800 to 4,000 Nvidia Grace Blackwell GPUs hosted entirely within Mistral's European data centers, underscoring the lab's push for infrastructure independence.

API Pricing Structure and Reported Benchmarks

On commercial pricing, Mistral lists standard API rates at $1.36 per 1 million input tokens and $4.18 per 1 million output tokens. Portions of Mistral's developer portal also display a temporary promotional rate of $0.68 per 1 million input tokens and $2.09 per 1 million output tokens to encourage early adoption during the preview period.

Mistral AI published initial performance metrics highlighting software engineering and automation strengths: 61.7% on the DeepSWE v1.1 benchmark, 59.9% on AutomationBench, and 82% on a reproduce-and-patch cybersecurity vulnerability evaluation. Independent evaluator Artificial Analysis separately registered an Intelligence Index score of 38 for the preview model.

Comparative Evaluations and Unresolved License Questions

While Mistral claims parity or advantages over leading open-weight architectures, independent third-party evaluations suggest that ML4 still trails top scores achieved by flagship models such as GLM-5.3 and Kimi K3 on aggregate reasoning benchmarks, indicating that raw scale does not automatically guarantee market dominance in multi-step problem solving.

Furthermore, the final open-weight license terms remain formally unpublished. Observers expect Mistral to release the weights under a custom corporate license rather than a permissive Apache 2.0 framework, leaving the exact enterprise distribution, self-hosting rights, and commercial restrictions unconfirmed until the public artifact release.

Practitioner Reactions and Deployment Friction

Practitioners and enterprise engineers broadly welcomed the announcement as evidence that European open-weight AI can construct frontier models capable of operating at the trillion-parameter scale alongside American and Chinese labs.

However, early practitioner sentiment was tempered by substantial practical concerns regarding self-hosting. Retaining 1.05 trillion parameters in memory requires extensive enterprise-grade hardware clusters, putting local on-premise execution out of reach for individual practitioners and small teams. Furthermore, several engineers attempting to integrate the preview API into agentic coding frameworks noted operational friction, finding that practical day-to-day tooling still faces integration hurdles compared to mature proprietary setups.

Strategic Implications for Enterprises in Thailand

For enterprise technology leaders in Thailand, Mistral Large 4 presents a viable alternative for processing massive textual inputs. The 1-million-token context window offers immediate utility for financial institutions, legal firms, and telecommunications operators seeking to analyze extensive compliance documentation, multi-year audits, and localized datasets across multiple languages without strict vendor lock-in.

Nevertheless, Thai enterprises pursuing data sovereignty through private on-premises hosting must weigh the significant capital expenditure required to serve a 1.05-trillion-parameter MoE model. Given the intense VRAM requirements, most regional organizations will likely find managed API consumption more economical during the preview phase, while monitoring Mistral's forthcoming licensing terms to verify compliance with local data governance standards before planning custom cluster deployments.