Company Updates

Mistral's $20M threshold splits open-weight AI models in 2026

Share
Mistral's $20M threshold splits open-weight AI models in 2026

Mistral and Moonshot AI have introduced revenue-based restrictions on their open-weight AI models, while DeepSeek and Alibaba offer more permissive licenses. This split is reshaping the open-weight vs. open-source debate in 2026.

TL;DR

  • Mistral and Moonshot AI impose revenue-based restrictions on their open-weight models, while DeepSeek and Alibaba offer more permissive licenses.
  • The split between open-weight and open-source models is becoming more pronounced, with implications for developers and startups.
  • Alibaba's Qwen3.8 is open-weight, but its flagship Qwen4-Max remains closed, complicating the landscape.

What happened

Mistral released the weights of Mistral Large 4 under a modified MIT license that requires a separate commercial deal for Model-as-a-Service businesses making over $20M annually. Moonshot AI's Kimi K3.1 Pro followed with a custom license, also imposing revenue-based restrictions. DeepSeek V4.1 Flash 1017 and Alibaba's Qwen3.8, on the other hand, are released under plain MIT and Apache 2.0 licenses, respectively, with no such restrictions.

The release dates for these models span from August to November 2026, with Mistral Large 4 and Kimi K3.1 Pro being the most recent. The models vary significantly in size, with Kimi K3.1 Pro having about 2.8 trillion parameters and DeepSeek V4.1 Flash 1017 having 552 billion parameters.

Why it matters

For developers and startups, the choice of model depends on the intended use case and revenue projections. Models with restrictive licenses may require additional legal agreements and could limit scalability for certain businesses. However, they may also offer advanced features and capabilities.

Investors should consider the licensing terms when evaluating AI startups. Companies using models with restrictive licenses may face higher legal costs and operational constraints, which could impact their growth and valuation. Conversely, startups using more permissive models may have greater flexibility and lower barriers to entry.

Key facts

  • Mistral Large 4: 1.05 trillion parameters, released October 29, 2026, modified MIT license with commercial use restrictions for Model-as-a-Service businesses making over $20M annually.
  • Moonshot AI's Kimi K3.1 Pro: About 2.8 trillion parameters, released November 4, 2026, custom license with revenue-based restrictions for hosted inference services.
  • DeepSeek V4.1 Flash 1017: 552 billion parameters, released October 17, 2026, MIT license with no revenue-based restrictions.
  • Alibaba's Qwen3.8: 2.4 trillion parameters, released August 17, 2026, Apache 2.0 license with no revenue-based restrictions. Qwen4-Max, the newer flagship, is closed and not open-weight.
  • All models support text and image input, with context windows of up to 1 million tokens.
  • Official API prices vary, with Kimi K3.1 Pro at $2/input and $10/output, Mistral Large 4 at $0.70/input and $2.10/output, and DeepSeek V4.1 Flash 1017 at $0.30/input and $1.20/output peak.

Context

The open-weight vs. open-source debate is becoming increasingly important in the AI industry. Open-source AI systems are defined by the four freedoms: use, study, modify, and share. Open-weight models, on the other hand, only provide the trained parameters and may have restrictive licenses.

The trend of imposing revenue-based restrictions on open-weight models is likely to continue, as companies seek to monetize their AI offerings while still providing access to developers. This could lead to a more fragmented market, with different models catering to different use cases and business models.

Topics

Related coverage

Join the discussion

Have a take on this story? Weigh in with our community on Facebook.

💬 Discuss on Facebook →