Anthropic's surprise move – suspending exports of its Mythos and Fable 5 models – instantly carved a gap in a strategic niche: advanced AI tools for security automation and complex workflow orchestration. Two Asian companies wasted no time and this week lifted the curtain on solutions designed to fill that exact hole. Tokyo-based startup Sakana AI introduced Fugu, an orchestration model that, according to initial benchmarks, holds its own against Fable 5. From Beijing, cybersecurity firm 360 Security unveiled Tulongfeng, a vulnerability-discovery tool aimed squarely at competing with Mythos.

Fugu and Tulongfeng: what we know

Technical details are scarce, but the positioning is crystal clear. Fugu isn't just another general-purpose LLM: it's a model built to orchestrate multiple AI components, coordinating complex tasks where reliability and low latency matter. Tulongfeng, for its part, is designed to unearth vulnerabilities in code and infrastructure – an area where generalist models often fall short.

Both launches are more than product announcements; they are political and market signals. The export ban imposed by Anthropic – tied to geopolitical tensions and U.S. restrictions on sensitive technology exports – has accelerated the hunt for homegrown Asian alternatives. And it opens a scenario familiar to anyone involved in on-premise AI: when access to cloud-first models disappears, the ability to run inference on your own infrastructure becomes a critical factor.

What this means for local deployment

For organizations already evaluating self-hosted stacks, the arrival of Fugu and Tulongfeng adds interesting building blocks. An orchestration model like Fugu, if truly competitive, could integrate with existing pipelines without depending on external APIs. A vulnerability-discovery tool like Tulongfeng, run on-premise, can probe air-gapped networks without exposing sensitive data.

Of course, heavy unknowns remain: we don't know VRAM requirements, context windows, FP16 precision, or real-world latencies. But the trend is unmistakable: the ecosystem is fragmenting, and data sovereignty is moving from slogan to procurement variable. Companies that built operations around Western models now face the unreliability of software supply chains.

The bigger picture: fragmentation and resilience

We're witnessing a catch-up race reminiscent of the CPU scramble after Huawei restrictions. Generative AI, until recently dominated by a handful of Californian vendors, is experiencing a forced diaspora: Japanese startups, Chinese security giants, and European initiatives are all trying to build independent stacks. The issue isn't just performance, but control over the model lifecycle – from quantization to fine-tuning, all the way to local serving.

For engineers deciding where to run critical workloads, the question is no longer 'cloud or not?', but 'which model can I run in-house without legal chains?'. AI-RADAR, which maps precisely these trade-offs for organizations choosing on-premise deployment, is watching the evolution of such alternatives closely. The ability to drop a tool like Tulongfeng into a local security pipeline, for example, reduces the attack surface and keeps logs under lock and key.

Outlook: a market going regional

Fugu and Tulongfeng are just the first examples of a trend that will see regional models multiply in the coming months. Anthropic's export ban created an opening, but the underlying demand – the ability to train and serve models without being tied to foreign jurisdictions – already existed and will only grow. Those investing today in on-premise infrastructure and open-source frameworks may find themselves with a competitive edge tomorrow, not just technological but legal as well.

In the meantime, eyes are on the first independent benchmarks and the spec sheets that Sakana AI and 360 Security will have to publish. Because if the numbers support the promises, the enterprise AI map could redraw itself faster than expected.