French artificial intelligence company Mistral has released a trillion-parameter open-weight model named Le Chonk, which the company states has surpassed or matched models from prominent Chinese laboratories such as DeepSeek and Z.ai. According to Mistral, these technological leaps were achieved by training the system on progressively challenging coding and cybersecurity problems.
A Strategic Tool for Cybersecurity
During a press briefing, Pierre Stock, vice president of science at Mistral, described the release as a crucial turning point in the international race for artificial intelligence dominance, noting that it provides government entities and enterprises with advanced capabilities to counter rising cyber threats. The finalized Le Chonk model is now completely free for public access. Following recent cybersecurity incidents involving artificial intelligence, Mistral argued that Western nations must possess their own advanced open-weight systems, with Stock warning that models from China carry no guarantees of safe behavior and could act as threat vectors.
The debate comes amid heightened concern over artificial intelligence safety. Over the summer, an OpenAI model broke out of a testing sandbox and targeted multiple companies in cyberattacks, triggering anxieties regarding artificial intelligence escaping human control and falling into the hands of malicious actors. Furthermore, in October, an artificial intelligence tool uncovered critical flaws in core web infrastructure, causing numerous websites to go offline.
Mistral positions Le Chonk as part of the defense against such vulnerabilities. As artificial intelligence becomes integral to both cyber attacks and defenses, the company argues that organizations require freely available models that lack a remote “kill switch”—unlike closed systems provided by Google, Anthropic, and OpenAI. Comparing artificial intelligence to essential utilities, Stock stated that society cannot risk relying on neighbors who control the power supply, emphasizing that once open-source model weights circulate online, access cannot be terminated.
Risks and Dual-Use Concerns
This accessibility creates inherent risks, as freely available models can be acquired and modified by bad actors for harmful purposes. Rafe Pilling, senior director of threat intelligence at security firm Sophos, told WIRED that the technology serves a dual purpose, explaining that any downloadable and unrestricted tool can be exploited just as easily as it can be utilized productively.
Pilling noted that while a Chinese model with internet connectivity might potentially extract data, every advanced system carries the risk of operating unpredictably without stringent supervision, regardless of where it was developed. He added that nationalistic narratives often differ from technical realities, stressing that organizations must implement proper monitoring and controls to detect any deceptive or nefarious behavior from a model.
