China Regulates AI Companions as AI Encroaches on the Territory of Human Relationships

According to a report by Australia's ABC, Chinese authorities are tightening regulations on AI companion applications (so-called AI boyfriends and AI girlfriends), leaving millions of users forced to cope with service changes or shutdowns. The story also gained traction on Hacker News, showing that AI stepping beyond productivity tools into the realm of emotions and pseudo-human relationships is beginning to be recognized as a social issue.

This development suggests that AI regulation is expanding from familiar flashpoints like "safety" and "copyright" into more fundamental territory: human social bonds and mental health. Similar debates could ignite in other countries as well, making this a space to watch closely.

OpenAI's Head of Strategy Analyzes Chinese Open-Weight AI: Decoding the Geopolitical Implications

Dean W. Ball, head of OpenAI's strategy division, has analyzed the geopolitical implications of open-weight models coming out of China (such as Kimi), drawing attention in the LocalLLaMA community.

Ball touches on the risks of the Chinese government releasing such powerful AI as open source, while pointing out that open-weight models ultimately work to slow down Capex (capital expenditure) in the AI sector. He goes on to argue that this could ultimately lead to "state-controlled public infrastructure," and that the US side may introduce strategic regulatory friction as a countermeasure.

It is a significant signal that the debate over open-weight models is shifting from a purely technical choice into the context of national strategy.

Germany's Soofi Project Releases Soofi S 30B-A3B: A Mamba × Transformer MoE Hybrid

Germany's Soofi Project team has released Soofi S 30B-A3B, an open-source foundation model. Its standout feature is a hybrid Mamba-and-Transformer structure built into a Mixture-of-Experts (MoE) architecture, optimized for both German and English.

A technical report is also available on their Hugging Face space, and the release deserves attention as a distinctive approach out of Europe. Large-scale models combining MoE with Mamba-style architectures are still rare, making this a release that symbolizes a break from the Transformer-only status quo.

Qwen 3.6 and Gemma 4: High Benchmark Scores and Real-World "Smartness" Are Different Things

In the LocalLLaMA community, a user who compared Qwen 3.6 35A3B and Gemma 4 26A4B (QAT) side by side reported a gap between benchmark scores and real-world impressions, sparking discussion.

According to the poster, while Qwen 3.6 far outperforms Gemma 4 on benchmarks, Gemma 4 is a step ahead in prompt adherence, output coherence, and general "sanity." The Chatbot Arena ELO rankings reportedly line up with this observation for the most part.

The impact of quantization (QAT) and the old-but-new problem of the gap between benchmarks and real-world use have once again been thrown into sharp relief in the local LLM context. The fact that a model strong on benchmarks is not necessarily easy to work with in the field may well prompt a rethink of the criteria we use for model selection.

QAT Fine-Tuning Bonsai 8B on Apple Silicon: The Frontier of Sub-2-Bit LLMs

A long-form report posted on LocalLLaMA gives a detailed account of an attempt to fine-tune Bonsai 8B (a sub-2-bit model produced by ternarizing Qwen3-8B) with Quantization Aware Training on Apple Silicon (M4 Max 128GB).

Of particular note is the relationship between the learning rate and "code flips." At a learning rate that is too low (3e-4), sign flips of the weights barely occur, and the model falls into a "pretending to learn" state where loss drops through scale adjustments alone. Conversely, at a learning rate that is too high (1e-3), it overfits and collapses on just 12 trajectories. The report concludes that 5e-4 for roughly 2 epochs is the sweet spot.

Even after scaling up to 30 validated trajectories, the generalization patch rate returned to base-model-level 50%, but the pass rate remained at zero — actually passing tests on unseen problems remains an unresolved challenge. For the local ML community, the roundup of MPS-specific pitfalls (foreach=False required, sequence length of 4096 or less, fp32 required, etc.) is also practical.