./daily — ← back to field notes

Daily links

One-line pointers to today's local-AI, model, hardware, and healthcare-AI news. No commentary — just the headline, the link, and where it came from.

Sunday, 2026-09-20·last updated 2026-09-20 23:18 UTC·9/9 feeds ok

[ hardware ]8

• b11065 — llama.cpp
• NVIDIA MMS1X00-N5400 QSFP112 1310NM 400Gbps Optic Quick Look — STH
• b11063 — llama.cpp
• b11062 — llama.cpp
• b11060 — llama.cpp
• STH Q3 2026 Letter from the Editor A Moving Target — STH
• b11056 — llama.cpp
• b11055 — llama.cpp

[ new models ]9

• DIY Jev — r/LL
• Would you buy a Qwen3.8-27B Taalas chip for $1k if it could run at 7,000 TPS? — r/LL
• I tested 9 LLMs on the exact same web-dev prompt for ~8 hours — RTX 3060 12GB results (R… — r/LL
• b11064 — llama.cpp
• Qwen-Image-2.1 released! — r/LL
• Reached 1.89 TB/s memory bandwidth overclocking the CMP 170hx — r/LL
• What are you all using for long term project/conversational memory these days? — r/LL
• b11058 — llama.cpp
• Your Agent Aced the Task. Will It Do It Again? — HF

[ local ai ]12

• Qwen3.8-Flash-Next Cosmic Arcade oneshot slop game — r/LL
• The bear can dance: Qwen 3.8 27B on one 3090 for 3 weeks — r/LL
• One more 'you should try ExllamaV3/exl3 for flash next' appreciation post — r/LL
• Lawsuit says Anthropic, OpenAI, SpaceXAI and Google made illegal agreement on AI slowdow… — r/LL
• A Jev-style model fine-tuned on Qwen3.5 4B — r/LL
• CUDA: enable sparse fa for qwen4 by am17an · Pull Request #28770 · ggml-org/llama.cpp — r/LL
• Is Typesafe based/derived from work done by the Laya author? — r/LL
• rene98c/Step-5-Preview-BF16 • HuggingFace (Fork) — r/LL
• What is JEV and what is it used for? — r/LL
• China's CXMT says new memory-chip platform enters mass production — r/LL
• this looks promising: stepfun-ai/Step-5-Preview-BF16 · Hugging Face — r/LL
• Finally got Qwen 3.8 Next running on my v100 6gpu setup (TP2 PP3) — r/LL

[ benchmarks & papers ]3

• Pirate Face Rescues LLM Models from Deletion — HN
• focus-llama: a llama.cpp fork implementing Declarative Attention (arXiv:2609.02737) — r/LL
• Cybernetics, interoception, and the art of embodiment — NatMI

[ tools ]4

• Speed-up Kimi K3(2.8T) on a 16x GB10 Cluster — 30 t/s coding throughput, 136 t/s concurr… — r/LL
• laya.cpp: Optimized laya near-instant decision making — r/LL
• b11059 — llama.cpp
• b11057 — llama.cpp

[ healthcare ai ]3

• Jehovah’s Witnesses allow blood-derived products but keep ban on whole-blood transfusion… — STAT
• Healthcare AI's next test is integration — HN-HC
• Inside The Medical AI Factory — HN-MED