The local LLM scene has seen a steady stream of activity around DeepSeek's new model this week. This time, a community-driven fix for broken tool calling has been merged into llama.cpp. Meanwhile, a fundamental question — how much further can model 'shrinking' go? — is stirring debate on LocalLLaMA. On the regulatory front, the injunction xAI sought against Minnesota's 'nudify' app ban was rejected. And signs of backlash and fatigue toward AI keep piling up, from browser choices to chess engine development, to a prominent YouTuber's confession.
DeepSeek V4 Flash tool calling fix merged into llama.cpp
DeepSeek V4 Flash (0731), released the other day, had been reported to loop or behave erratically when using tool calling (function calling) in local environments. A fix for this issue was reportedly merged into llama.cpp, the major inference engine, about half a day ago. After the fix PR (#26269) landed in the mainline, the original reporter says the problems have not recurred on their end. It's a practically valuable update for those experimenting with agentic use cases locally, and it stands as a continuation of the ongoing DeepSeek V4 Flash story.
How small can models get?
Prompted by DeepSeek V4 Flash's smaller footprint and stronger performance, LocalLLaMA debated the question of 'how small can models get?' Thanks to distillation, MoE, synthetic data, and longer inference time, models with the same parameter count are indeed smarter than they were a few years ago. But the crux of the question is whether a 'minimum capacity' required for language understanding, broad knowledge retention, reasoning, and generalization exists. Some take a cautious view: even when benchmark scores approach those of large models, small models may still crumble on rare knowledge or unexpected prompts. Another interesting observation: the cost hasn't disappeared — it may have simply moved elsewhere, into expensive training runs and distillation from larger models.
Predicting Qwen's next open-source move
A user who has been happily running Qwen 3.6 35B-A3B posted speculation about what comes next for Qwen's open-weight releases. While Qwen 3.7 is already available via API through OpenRouter, whether its weights will be published remains unclear. The community voiced expectations in several directions, from the release of larger models to an expanded edge-focused lineup. Note that nothing here is confirmed — it remains pure community speculation.
Minnesota judge rejects xAI's bid to block 'nudify' app regulation
In the lawsuit in which xAI sought an injunction against Minnesota's state law banning 'nudify' apps that generate nude images from photos, the judge rejected xAI's request. This clears the way for the law to move forward. The pattern of AI companies challenging regulation on free-speech grounds — alongside the social problems of deepfakes and nonconsensual nude generation — is likely to repeat in other places going forward. The facts here are as reported by the press, and the ultimate judicial outcome remains to be seen.
AI backlash and fatigue ripple through communities
What's interesting is how many overlapping voices signal backlash and fatigue toward AI itself. On Hacker News, a post searching for 'a browser without AI features' drew widespread sympathy, with commenters arguing that the presence or absence of AI now shapes a piece of software's philosophy and its stance on privacy. In the chess engine development community, openly hostile reactions to AI-assisted development have reportedly surfaced. And prominent YouTuber Hank Green candidly confessed that the amount of dopamine he gets from interacting with LLMs is 'not good for me or for the world.' Each of these is an isolated event on its own, but taken together they suggest that as AI permeates everything, a desire to keep one's distance is spreading even into corners of the tech community.