What happened
Mistral AI published the complete model weights for Mistral Large 4, a 1‑trillion‑ , on Hugging Face and announced usage pricing and licensing terms.
On October 29, 2024, Mistral AI announced that the weights for its Mistral Large 4 (ML4) model are publicly available via Hugging Face (https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B). The model, codenamed le Chonk, contains 1 trillion total parameters and 49 billion activated parameters (52 billion when counting embedding and output layers). Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centre.
The release follows an earlier announcement three weeks prior that introduced the model architecture. Mistral notes that the current weights have been "optimized and upgraded" after real‑world testing with government and enterprise partners, suggesting improvements in stability and performance over the initial preview.
Pricing for online usage is disclosed: $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. The company also provides a score of 48 % on the Artificial Analysis Coding Agent Index, close to the leading open‑source GLM‑5.3 model, while claiming a 47 % reduction in average tokens consumed per task.
Mistral distributes the model under a modified MIT license. The license permits free personal, academic, modification, fine‑tuning, and general commercial use. However, cloud service providers or vendors that intend to offer the model as a managed service (MaaS) must obtain a separate commercial license from Mistral before resale.
Why it matters
The release marks one of the largest open‑weight language models to date, lowering the barrier for researchers and developers to experiment with trillion‑scale AI. By providing transparent pricing and a permissive license for most use‑cases, Mistral aims to accelerate adoption while retaining control over large‑scale commercial services. The model’s reported efficiency gains—nearly halving token consumption compared with comparable open‑source models—could reduce inference costs and broaden practical deployment in enterprise settings.
The open‑weight release of a trillion‑ model is a significant milestone in the AI ecosystem. Historically, models of this scale have been kept behind proprietary APIs due to the massive compute and data requirements needed for training. By open‑sourcing the weights, Mistral enables the research community to study scaling behaviours, develop novel fine‑tuning techniques, and against other large models without incurring the prohibitive cost of training from scratch.
Mistral's pricing disclosure is unusual for an open‑weight model. While the weights are free to download, the company monetises inference through a token‑based fee structure. This hybrid approach could set a precedent for sustainable business models that balance openness with revenue generation, especially as inference costs for trillion‑scale models remain high.
The modified MIT license draws a clear line between unrestricted personal/academic use and commercial MaaS offerings. By requiring a separate commercial license for resale, Mistral retains leverage over large‑scale deployments that could otherwise undercut its own cloud services. This licensing nuance may influence how other open‑weight projects handle commercial exploitation, potentially prompting more nuanced open‑source licenses in the AI domain.
Performance claims—particularly the 47 % token‑efficiency improvement—suggest that Mistral has focused on architectural or training optimisations that reduce compute per token. If validated independently, such efficiency could lower operational costs for downstream applications, making trillion‑ capabilities more accessible to smaller organisations.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
What to watch next
Watch for adoption trends among cloud providers, the emergence of downstream fine‑tuned variants, and any policy discussions around licensing of high‑capacity open models.
Adoption by cloud platforms: Major providers (e.g., AWS, Azure, GCP) may integrate ML4 into their AI marketplaces. Monitoring pricing parity and licensing negotiations will reveal how Mistral's commercial‑license model scales.
Community fine‑tuning: Open‑source contributors are likely to produce specialised variants (e.g., instruction‑tuned, domain‑specific). The quality and safety of these derivatives will be a key indicator of the model's practical impact.
Regulatory scrutiny: As governments evaluate the risks of large open models, Mistral's licensing terms and the model's multimodal capabilities could attract policy attention, especially regarding export controls or misuse mitigation.
verification: Independent labs may test the reported token‑efficiency and AA‑CAI Index score. Discrepancies could affect the model's credibility and influence future adoption decisions.