Sakana AI strips censorship from frontier models with Namazu

Top-tier open-source LLMs are ridiculously capable, but they come with a known catch: they tend to inherit the geopolitical biases and censorship reflexes of wherever they were trained. If you ask a frontier Chinese or US model about sensitive historical or diplomatic topics, you often get a heavy spin or an outright refusal.

Tokyo-based Sakana AI is tackling this head-on. The lab just released Sakana Chat alongside a new prototype model family called Namazu (alpha). Instead of spending hundreds of millions building frontier foundation models from scratch, Sakana took the world's best open-weight LLMs and used surgical post-training to align them for Japanese users.

Dropping the refusal rate to zero

The Namazu lineup includes three adapted variants: Namazu-DeepSeek-V3.1-Terminus, Llama-3.1-Namazu-405B, and Namazu-gpt-oss-120B. Sakana noted that the naming for the 405B version specifically follows Meta's licensing requirements.

The most striking benchmark is censorship removal. According to Sakana's evaluations on sensitive political, historical, and diplomatic queries involving Japan, the base DeepSeek-V3.1-Terminus refused to answer a whopping 72% of the time. After Sakana's post-training, that refusal rate dropped to nearly 0%, with substantial improvements in factual accuracy and multi-perspective neutrality.

Crucially, modifying the models didn't lobotomize their core smarts. Across demanding reasoning and technical benchmarks like AIME'25, GPQA Diamond, MMLU-Redux, and LiveCodeBench, Namazu models preserved parity with their base architectures. On Japanese benchmarks like Nejumi Leaderboard4 and Swallow, the DeepSeek variant matched the original base model and comparable domestic rivals.

To make these accessible, Sakana embedded them into Sakana Chat with native real-time web search. The models were trained using compute from GMO Internet's GMO GPU Cloud between October and November 2025, following a private beta with about 1,000 testers. Weights and a technical paper are coming soon.

Why it matters

For users in Japan, this delivers a chat interface powered by 405B-class frontier reasoning that actually discusses delicate regional history and current affairs objectively, rather than dodging questions. Having real-time web search baked in makes it an immediate practical alternative for researchers, journalists, and knowledge workers.

For the wider AI race, Sakana is charting the only sensible playbook for non-superpower nations. Frontier pre-training is rapidly becoming a US-China duopoly due to crushing capital requirements. Trying to build a sovereign base model from the ground up in Europe or Japan is often an expensive exercise in reinventing the wheel. Taking the world's best open weights and mastering cultural post-training is vastly cheaper, faster, and more effective at delivering sovereign AI.

The strategic risk, however, is dependency. You are tethered to the goodwill and licensing terms of foreign model builders. Furthermore, aggressively stripping away base-model guardrails to lower a 72% refusal rate to near zero is technically impressive, but it creates a delicate balancing act between removing foreign censorship and maintaining essential safety protections once those weights are made public.