An open-weight AI model just broke a record that only held for a week. On July 16, 2026, Chinese start-up Moonshot AI unveiled Kimi K3, a 2.8 trillion-parameter model, the largest ever released in open access. For an SME leader comparing ChatGPT, Claude, or Copilot, this announcement is not just a technical feat: it reopens the question of cost, single-vendor dependency, and, since this is a Chinese model, data sovereignty.
In brief
- Moonshot AI (China) announced Kimi K3 on July 16, 2026: 2.8 trillion total parameters, a one-million-token context window, available immediately via the Kimi API.
- The full model weights are due to be released on July 27, 2026, under a modified MIT license, making self-hosting possible instead of depending on a single vendor.
- According to independent evaluations by Artificial Analysis, K3 outperforms Claude Opus 4.8 and GPT-5.5 on several coding and agentic benchmarks, while still trailing the newest models such as Claude Fable 5 or GPT-5.6 Sol.
- Pricing: $3 per million input tokens, $15 per million output tokens, roughly three times the price of Moonshot AI's previous model, yet still lower than comparable closed models.
- For an SME, the point is not to adopt Kimi K3 tomorrow, but to understand what the rise of open-weight models changes in the balance of power with major US vendors.
Kimi K3, explained
Kimi K3 is a large language model developed by Moonshot AI, a Chinese AI lab already known for its Kimi model family. It uses a "mixture of experts" architecture: out of 896 specialized sub-networks, only 16 are activated per request, which limits the compute needed despite the model's total size. Its one-million-token context window lets you feed it an entire codebase or several hundred pages in a single request.
One important distinction: Kimi K3 is released in open weights, not "open source" in the strict sense. The trained model files will be downloadable and modifiable under a modified MIT license, but the training data and code remain Moonshot AI's property. It's the same logic as Llama (Meta), Mistral, or DeepSeek: a company can host the model on its own servers, without depending on the vendor's API.
A race that keeps accelerating
Kimi K3 succeeds the Kimi K2.6 model and is roughly 75% larger than DeepSeek V4 Pro, which previously held the record for the largest Chinese open-weight model, according to figures published by Moonshot AI. The release timeline shows just how fast this race is moving.
July 16, 2026
Kimi K3 announced
July 27, 2026
Open weights released
August 2, 2026
EU AI Act comes fully into force
Several business media outlets compared this announcement to the shock caused by DeepSeek in early 2025: a signal that China can still produce world-class models, at training and usage costs lower than those of US labs.
Open weights vs. proprietary models: what it changes for an SME
The vast majority of French SMEs today use proprietary models, ChatGPT, Claude, or Copilot, through a simple subscription. An open-weight model like Kimi K3 opens a different option, with its own trade-offs.
Proprietary model (ChatGPT, Claude, Copilot)
Access only through the vendor's API or interface. Pricing, terms, and updates decided by a single player. No data leaves the vendor's servers, but no control either over exactly where it is located.
Open-weight model (Kimi K3, Llama, Mistral)
Downloadable model files, hostable with a European provider or in-house. More control over data location and long-term cost, but more technical responsibility for setup and maintenance.
This choice connects to a question already raised on Le Mag LUWAI about the risk of depending on a single AI vendor: the more a company builds its processes around a closed tool, the more expensive it becomes to switch. The rise of open-weight models, even if not yet suited to every SME, widens the options available to avoid that dependency.
What to check before adopting a Chinese open-weight model
Technical enthusiasm shouldn't overshadow the practical questions an SME needs to ask before installing or using a model like Kimi K3.
| What to check | Why it matters |
|---|---|
| Hosting location | Using the Kimi API means data transits through servers located outside the European Union, with a GDPR framework to clarify |
| Available technical skills | Self-hosting a model of this size requires infrastructure and expertise few SMEs have in-house |
| Support and maintenance | An open-weight model doesn't come with the customer support of a standard subscription |
| Vendor security maturity | The Future of Life Institute ranked Chinese AI labs among the lowest-scoring in its 2026 safety index |
Key takeaway
For the vast majority of SMEs, there's no urgency to switch AI tools in reaction to Kimi K3. The short-term relevance lies elsewhere: this announcement puts pressure on pricing and pushes proprietary vendors to justify their added value, a balance of power that benefits any business buying AI.
FAQ
What is an open-weight AI model?
It's a model whose trained parameters are published and downloadable, making it possible to install on your own servers. Unlike a fully "open source" model, the training data and code usually remain private.
Is Kimi K3 more capable than Claude or GPT-5.6?
According to independent evaluations by Artificial Analysis, Kimi K3 outperforms Claude Opus 4.8 and GPT-5.5 on several coding and agentic benchmarks, but still trails the newest models such as Claude Fable 5 or GPT-5.6 Sol on the most demanding evaluations.
Can a French SME use Kimi K3 safely?
Through the Kimi API, data transits through servers located outside the European Union, which requires checking the processing terms before any use involving sensitive data. Self-hosting the open weights, expected from July 27, 2026, would let data stay in-house, but requires technical skills few SMEs have internally.
Should you wait for the full weights release before acting?
For most SMEs, yes: there's no urgency to switch tools. That said, it's a good moment to watch whether your current AI vendor adjusts its pricing or terms under this increased competition.
In summary
Kimi K3 won't replace ChatGPT or Claude in an SME's daily use overnight. But the announcement confirms a deeper trend: competition between proprietary and open-weight models is intensifying, and it benefits companies that keep an eye on their options rather than locking themselves into a single tool. To go further on choosing and governing your AI tools, browse the other guides on Le Mag LUWAI or see how our clients structure their AI usage in our success stories.
