Mistral, the French AI lab, announced its latest model, “Le Chonk,” now matches or outperforms DeepSeek’s best models in coding and cybersecurity, according to Wired. This competitive claim surfaces as DeepSeek faces intensifying regulatory pressure from the United States government.
DeepSeek has kept up a rapid pace of releases, even as geopolitical divisions grow. This outlet first covered DeepSeek’s V4 Pro launch on August 16, noting its unexpected performance. The company has since released new tools and models. On October 4, DeepSeek Harness v0.2 launched, bringing official desktop applications to its open-source agent harness. Less than two weeks later, on October 18, DeepSeek released V4.1 Flash, which scored 77.0 on the DeepSWE benchmark and opened a search API. Just three days later, on October 21, DeepInfer, an open-source inference framework, also became available. This quick succession of launches points to the company’s focus on expanding its AI offerings and developer tools.
These technical advancements and product rollouts happen alongside significant regulatory challenges. Starting October 20, the US government moved to ban DeepSeek from federal civilian agency IT systems. This directive, first reported by The Hindu and also covered by The Economic Times, cited national security concerns. Further reports from Startuptalky on October 21 and 22, and Analytics India Magazine on October 22, confirmed the order for DeepSeek’s removal from federal systems. Rapid AI development and national security policies now face increasing friction. The overlap of these actions with DeepSeek’s model launches demonstrates this.
Mistral’s claim about “Le Chonk” directly addresses the “open-weight race,” a competition among AI developers to release the most capable models with publicly available weights. DeepSeek participated prominently. Its models like V4.1 Flash demonstrated strong benchmark performance. Mistral’s assertion that its model now rivals or exceeds DeepSeek and Z.ai in specific areas like cybersecurity directly challenges the perceived lead of certain Chinese AI developers within the open-source community. This model represents a declared intent to compete at the very top of the open-source capability ladder.
Prime Intellect released Prime Mega MoE, a backward pass for DeepSeek’s fused Mega MoE kernel. This development, reported by MarkTechPost, makes the kernel trainable and offers significant performance gains. Prime Mega MoE speeds Qwen3 MoE training steps by up to 34% in prime-rl, with a 3.11x faster forward pass. DeepSeek’s underlying architectural contributions are thus finding adoption and improvement within the wider AI research community, even as the company faces external competitive and regulatory pressures.
DeepSeek releases models quickly and contributes foundational technology, such as the Mega MoE kernel, now made trainable by Prime Intellect. At the same time, it confronts direct competitive challenges, including Mistral’s “Le Chonk” claiming parity, and significant geopolitical barriers, shown by the US ban from federal civilian systems. The V4.1 Flash model scored 77.0 on DeepSWE. This score confirms its technical capabilities remain a point of reference for competitors.
What is the significance of Mistral’s “Le Chonk” model?
Mistral claims its “Le Chonk” model matches or outperforms DeepSeek’s best models in coding and cybersecurity.
What regulatory action has the US taken against DeepSeek?
The US government ordered the removal of DeepSeek from federal civilian agency IT systems starting October 20, citing national security concerns.
What technical contribution did Prime Intellect make related to DeepSeek?
Prime Intellect released Prime Mega MoE, a backward pass for DeepSeek’s fused Mega MoE kernel, making it trainable and speeding Qwen3 MoE training steps by up to 34%.
Related Coverage
Compiled by Launch91 Desk from the sources linked above. More about Launch91.