IBM, Mistral AI, and the Technology Innovation Institute (TII) have released three open-weight language models in the past 16 months, each targeting enterprise and government customers prioritizing data sovereignty over benchmark scores. IBM Granite 4.2, Mistral Large 3, and Falcon H1 offer distinct advantages in compliance, multilingual support, and cost efficiency.
TL;DR
- IBM Granite 4.2, Mistral Large 3, and Falcon H1 are open-weight models designed for enterprise and government use, with a focus on data sovereignty and compliance.
- Granite 4.2 offers strong reasoning capabilities and is built for auditable enterprise agents, while Mistral Large 3 provides a large context window and multimodal support, and Falcon H1 specializes in Arabic language tasks and offers a range of model sizes.
- The models compete on parameters, context windows, pricing, and licensing, with each offering unique advantages for specific use cases.
What happened
IBM released Granite 4.2 on August 25, 2026, with three model sizes (3B, 8B, and 30B parameters) and a focus on enterprise agents that require an audit trail. The model is designed to be open, performant, and trusted, with support for a 128K-token context window and a 512K-token window for the 30B model. Granite 4.2 is released under the Apache 2.0 license and includes cryptographic signing and ISO certification. IBM's benchmark table shows strong reasoning capabilities, with the 30B model scoring 89.17 on AIME25 and 77.60 on MMLU-Pro (five-shot).
Mistral AI released Large 3 on December 2, 2025, with a 675-billion-parameter mixture-of-experts model and a 256K-token context window. The model is multimodal, supporting text and image inputs, and is released under the Apache 2.0 license. Mistral's API, La Plateforme, offers competitive pricing at $0.50 per million input tokens and $1.50 per million output tokens. Mistral has raised significant investment, with a valuation topping €21 billion after a €3 billion investment round involving Samsung. The model is positioned as Europe's answer to US and Chinese frontier labs, with a focus on technological autonomy and data sovereignty.
TII released Falcon H1 on May 21, 2025, with a family of six models ranging from 500M to 34B parameters, including an Arabic-first variant and a 7B reasoning model, Falcon H1R, released in January 2026. The models are released under the TII Falcon License, which is Apache-2.0-based with an added acceptable-use policy. Falcon H1 offers strong Arabic language capabilities, with the 34B model scoring 75.36% on the OALL leaderboard average for Arabic-specific evaluation. The models are designed for self-hosting, with no official hosted API pricing published as of this writing.
Why it matters
For developers and startups, these models offer open-weight alternatives to proprietary models, with a focus on data sovereignty and compliance. Granite 4.2 is ideal for enterprise agents that require an audit trail, while Mistral Large 3 offers a large context window and multimodal support. Falcon H1 specializes in Arabic language tasks and offers a range of model sizes for different use cases. The models compete on parameters, context windows, pricing, and licensing, with each offering unique advantages for specific use cases.
For investors, the models represent a shift in the AI market towards open-weight models and data sovereignty. Mistral's significant investment and valuation highlight the growing interest in European AI infrastructure. The models also raise important questions about the tradeoffs between open-source AI and security, as highlighted by Palo Alto Networks' warning about the risk profile of open-source AI. As the AI market continues to evolve, these models are likely to play an increasingly important role in enterprise and government deployments.
Key facts
- IBM Granite 4.2 was released on August 25, 2026, with three model sizes (3B, 8B, and 30B parameters).
- Mistral Large 3 was released on December 2, 2025, with a 675-billion-parameter mixture-of-experts model and a 256K-token context window.
- Falcon H1 was released on May 21, 2025, with a family of six models ranging from 500M to 34B parameters, including an Arabic-first variant and a 7B reasoning model, Falcon H1R, released in January 2026.
- Granite 4.2 is released under the Apache 2.0 license and includes cryptographic signing and ISO certification.
- Mistral Large 3 is released under the Apache 2.0 license and offers competitive pricing at $0.50 per million input tokens and $1.50 per million output tokens.
- Falcon H1 is released under the TII Falcon License, which is Apache-2.0-based with an added acceptable-use policy.
- Granite 4.2 offers strong reasoning capabilities, with the 30B model scoring 89.17 on AIME25 and 77.60 on MMLU-Pro (five-shot).
- Mistral Large 3 offers a large context window and multimodal support, with a focus on European technological autonomy and data sovereignty.
Context
The release of these models highlights a growing trend in the AI market towards open-weight models and data sovereignty. As governments and enterprises increasingly prioritize control over their data and infrastructure, these models offer an alternative to proprietary models from US and Chinese companies. The models also raise important questions about the tradeoffs between open-source AI and security, as well as the role of AI in promoting technological autonomy and independence. As the AI market continues to evolve, these models are likely to play an increasingly important role in enterprise and government deployments, as well as in the broader AI ecosystem.
