The M5 Ultra Mac Studio is presented as a practical breakthrough for running powerful local AI agents. Its UltraFusion quad-die design pairs two dual-die M5 Max chips into an 80-core GPU with Neural Accelerators and raises unified memory bandwidth from 819 GB/s to 1.2 TB/s while retaining a 512 GB memory ceiling. In hands-on tests against an M3 Ultra (512 GB) and an RTX 5090 desktop, the M5 Ultra delivers substantially faster prompt processing and token generation: about 70% faster response generation on average versus the M3 Ultra and roughly a 150% improvement in prompt processing (≈2.5×). That lets models like Qwen3.8-Flash-Next and Qwen-Image-2.1 run responsively in agent frameworks such as Open Minis, Hermes, and Codex, sustaining 100 tokens/sec on short prompts and 60-85 tokens/sec with very large contexts (64K-256K), enabling fluid multi-turn, tool-using workflows.
Those performance gains translate into concrete productivity: a custom research app called Desk used a team of local agents 24/7 for 99 days to transcribe WWDC sessions, extract and cross-reference features from 310 documents, OCR PDFs, and sync results via the Notion API, all at zero cloud cost. The machine’s thermal and acoustic profile and macOS ecosystem make it preferable to a noisy high-end PC for this use, though a discrete GPU still has advantages in raw bandwidth. This setup is powerful but fiddly - ideal for tinkerers or privacy- and cost-conscious users who want always-on local agents, not for those who prefer turnkey cloud services.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Open language models include open-weight and open-source types, with Chinese companies leading in open-weight models since 2025. The competition between American and Chinese models influences the global AI landscape and its commercial viability. (interconnects.ai)
No Sloptober challenges individuals to abstain from using LLM-based tools throughout October to develop personal skills and awareness of AI's limitations. Participants are encouraged to reflect on their reliance on AI, improve their coding abilities, and reassess the value of automation in their work and learning processes. (no-sloptober.com)
Open-weight AI models now process 56% of tokens in production, up from less than 10% in December 2025. However, proprietary systems still generate most of the revenue, as they cost significantly more per inference. (techstrong.ai)
LLM ASSBENCH is a platform that evaluates large language models (LLMs). It currently shows no matches or data for prompts entered. (assbench.com)
Fine-tuning small language models is discussed, with users encouraged to share their tasks, models, and outcomes. The post seeks practical experiences and results from those who have tried this approach. (news.ycombinator.com)
Vals' team used AI agents to develop and formally verify a new shortest path algorithm called C-HD, which improves efficiency for directed graphs with non-negative weights. The algorithm achieves a new complexity bound, enhancing performance over previous methods. (vals.ai)
Today's best Hacker News stories, summarized and screenshotted, one email a day.