./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.

Friday, 2026-09-25·last updated 2026-09-25 23:15 UTC·9/9 feeds ok

[ hardware ]12

• b11191 — llama.cpp
• b11190 — llama.cpp
• I ran the actual break-even math on buying vs renting an H200 box, and it is not where I… — r/LL
• b11189 — llama.cpp
• b11188 — llama.cpp
• b11185 — llama.cpp
• Trained my first small language model — r/LL
• b11184 — llama.cpp
• b11183 — llama.cpp
• AMD Takes the Lid off of Next-Gen EPYC 9006 Venice As Zen 6 Comes to Servers — STH
• b11182 — llama.cpp
• b11181 — llama.cpp

[ new models ]7

• How long can I expect to wait until the local ~30B A3B frontier catches up to GLM 5.3 Fl… — r/LL
• 4-5 days replacing Claude w Qwen 3.8 Next — r/LL
• Qwen3.8-27B: Using KV Cache Transplants to Boost Output Quality — r/LL
• Swift1.5-Qwen3.8-Flash-Next is phenomenal vs. base 3.8-Flash! — r/LL
• Splash on a 40-core M5 Max: +20% decode by tuning the kernels for your own chip — r/LL
• Is Qwen Flash Next at like Q2 better than 27B at Q4? — r/LL
• Accelerating vision-language models with LFM2.5-VL-DSpark — HF

[ local ai ]12

• Mica v0.1 4B got an iron pickaxe in real Minecraft without generating a single token — r/LL
• Is there a lightweight version of Hermes agent? — r/LL
• Ion v0.2.0 — No install. No backend. Just one HTML file. — r/LL
• Make Volta Fast Again — r/LL
• How do you use subagents & multiple agent with local models, and how many? — r/LL
• PSA for M5Ultra owners running LLMs: set your prefill step to 8192 — r/LL
• Has anyone benchmarked AI agents against the SOLIDWORKS CSWA exam? — r/LL
• Gemma 4 Developer Agent Competition — r/LL
• Former Intel CEO: "HBM is lousy". High Bandwidth Flash Is Coming — r/LL
• How are you guys thinking about context now, and building around it? — r/LL
• I added OpenVINO support to Laya: 40 ms per question on CPU, 3.4x faster than PyTorch — r/LL
• Asianometry: High Bandwidth Flash: What Is It Good For? — r/LL

[ benchmarks & papers ]3

• Jev vs. Kev: open-source Jev alternative tested side by side — r/LL
• Qwengram-0.8B: I transferred Qwen3.8 Flash-Next’s n-gram memory into Qwen3.5-0.8B — 5.05… — r/LL
• Guiding large language models to predict edit sequences for molecular synthesizability o… — NatMI

[ tools ]1

• Best LLM for every budget, updated daily — HN

[ healthcare ai ]12

• STAT+: Licensing deals on generic versions of Roche flu drug aimed at preparing for pand… — STAT
• STAT+: Lilly wins approval for once-weekly insulin shot — STAT
• Opinion: FDA commissioner nominee Heidi Overton is ignoring the science on abortion drug — STAT
• STAT+: Up and down the ladder: The latest comings and goings — STAT
• STAT+: Pharmalittle: We’re reading about a reticent FDA commish nominee, a Novo deal, an… — STAT
• Federal judge rolls back landmark disability protections — STAT
• How Big Tobacco flooded the FDA with public comments to kill a menthol cigarette ban — STAT
• At least five candidates vie to become the next WHO leader, as questions over U.S. role… — STAT
• STAT+: The head of Kennedy’s new autism project has roots in Iraq War, Sesame Street — STAT
• Opinion: RFK Jr.’s use of the National Center for Health Statistics will add confusion t… — STAT
• Opinion: The ‘died with’ vs. ‘died of’ measles discussion misses something very importan… — STAT
• STAT+: FDA commissioner nominee Heidi Overton keeps opinions to herself in Senate hearin… — STAT