DeepSeek V4 Flash Gets First llama.cpp Support — A New Era for Local Inference?

A PR adding DeepSeek V4 series support to llama.cpp has been published on GitHub. The PR is still in its early stages, with a slow runtime of 5–6 tps, but accuracy is reportedly already at a verifiable level.

When a community user actually quantized and tested DeepSeek V4 Flash, the verdict was that it is the first local model in its size class to feel like it has "intelligence rivaling frontier models." Three points are particularly noteworthy. First, its native FP4-FP8 hybrid design makes it highly resilient to quantization. Second, its KV cache context scaling is extremely efficient. Third, while competitors such as MiniMax M2.7 struggle with quantization, DeepSeek V4 Flash maintains solid performance even at Q3 quantization.

The user also noted that "the Qwen 3.5/3.6 series is excellent, but DeepSeek V4 Flash goes a step beyond," predicting that it will hold a dominant position in the 80–140GB model class for several months to come.

MIT Tech Review: Are AI Chatbots Taking Control of Our Brains?

MIT Technology Review has published an article asking whether AI chatbots are taking control of our brains. It raises concrete concerns about how everyday conversations with AI affect human cognitive processes.

The article offers a detailed analysis and has already sparked discussion on Hacker News, picking up 3 points. The balance between the convenience of AI tools and cognitive dependency will likely remain a topic of significant social importance going forward.

Pearl's "Useful PoW" AI Mining Exposed as Vaporware

An accusatory post has surfaced on Hacker News claiming that Pearl's "useful" PoW AI mining is vaporware. The poster argues that although Pearl touts a useful Proof of Work leveraging AI computation, miners actually perform no inference — it is simply plain PoW wrapped in an AI facade.

Circumstantial evidence of involvement by Together AI and NVIDIA has also been presented. According to the post, Together AI hosts LLM inference on its own GPUs and subsidizes it with Pearl ($PRL) profits, yet miners on the Pearl network compute wasteful PoW instead of performing actual inference. Concern is growing in the community that this wasteful computation is driving GPU prices even higher.

Opus Dominates Low-Level Systems Engineering; the Gap Between Local and Frontier Models Remains Wide

An interesting post comparing local and frontier models on low-level systems engineering ability is drawing attention in Reddit's LocalLLaMA community.

Having tried various models on an AirPlay speaker firmware modification project, the poster reported that among frontier models including GPT-5, only Claude Opus succeeded in accurately mapping the firmware layout and reverse-engineering the CRC structure. The local model Qwen 3.6 35B-A3B is excellent for everyday use, but for this kind of task the gap with frontier models remains large, according to the post.

The community has also shared benchmark comparisons of the latest local models (sized to run on 3×3090s), and results for the Gemma-4 12B are eagerly awaited.

Japan's Developer Community: Claude Code Articles Continue to Thrive

On Japan's Qiita, technical articles about Claude Code continue to make the trending list.

Switching models via Amazon Bedrock**: A walkthrough of the correct configuration steps, warning that mistakes in model-switching settings when using Claude Code through Bedrock could lead to security incidents.

A coder ships an app in 3 weeks with Claude Code**: An account from an engineer who had been away from coding, learned Claude Code, and released an app in just three weeks.

Claude Code × MCP × Google Sheets integration**: Five patterns for automating Google Sheets operations by combining Claude Code with MCP.

Multi-layered gates to block emojis**: A proposed method for mechanically blocking Claude Code's unsolicited emoji output at three layers: pre-commit, review, and deploy.