./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, 2026-10-25·last updated 2026-10-25 11:58 EDT·9/9 feeds ok

[ hardware ]12

• b11878 — llama.cpp
• b11877 — llama.cpp
• b11876 — llama.cpp
• b11875 — llama.cpp
• b11874 — llama.cpp
• b11873 — llama.cpp
• With the huge price hike on the 32GB VRAM scene, I’m reconsidering my purchase options.… — r/LL
• Lean-QwenFlash: Qwen3.8-Flash-Next (125B) running at ~20 tok/s on 16gb vram and 8gb ram… — r/LL
• b11871 — llama.cpp
• b11870 — llama.cpp
• CMP170HX inspiration unlocks 30x/40x also, up to 2X Prefill speedup --> NVIDIA RTX 30 &… — r/LL
• b11869 — llama.cpp

[ new models ]6

• Nanbeige/Nanbeige4.3-33B · Hugging Face — r/LL
• CUDA: fuse FFN gate/up matmuls and GLU in MMQ with interleaved gate/up rows by anujj · P… — r/LL
• b11879 — llama.cpp
• OpenAI and Anthropic blocked the reverse engineering of smart TVs spying on us — r/LL
• Are there any good finetunes for Blender and/or 3D modeling? — r/LL
• LitPivot: Developing research ideas through dynamic contextualization within the literat… — HF

[ local ai ]12

• Backup your quants! — r/LL
• About running deepseek 4.1 flash — r/LL
• Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence — r/LL
• A 4B coding model that can generate 3D scenes — r/LL
• Too good to be true? ISTA's GSQ-RCO 2-bit DeepSeek-V4.1-Flash scoring the same as the ba… — r/LL
• GPT-X3-150M takes #1 on the Open SLM Leaderboard, dethroning HuggingFace's SmolLM2-135M… — r/LL
• model: DSV4.1 flash by am17an · Pull Request #31086 · ggml-org/llama.cpp — r/LL
• Is it just me or Kimi K3.1 Pro don’t look much better than DeepSeek 4.1 Flash? — r/LL
• UK to invest £100m in 'world first' AI Biosecurity Lab — r/LL
• Small model good for reverse engineering that can run on a 3090 and 32GB RAM. — r/LL
• High-Performance llama.cpp fork for Legacy GPUs: Kepler Support, CUDA Graphs & Windows N… — r/LL
• Made a rhythm game where 4 small neural nets listen to your song right in the browser ta… — r/LL

[ benchmarks & papers ]5

• [Benchmark] Strata/Basalt runs Qwen3.8 Flash Next IQ3_S with 16GB VRAM and 55GB RAM with… — r/LL
• M6 Mac Mini owners (32GB): now you can run Qwen3.8-27B in high quality on the Apple Neur… — r/LL
• KLD is misunderstood — r/LL
• Guidelines for evaluating structure-based methods for protein–ligand binding prediction — NatMI
• Efficient equivariant framework for universal charge density prediction — NatMI

[ tools ]1

• Using LLM for confidence scoring is a bad idea — HN

[ healthcare ai ]6

• STAT+: After a cancer vaccine’s first success, the race is on to expand the impact — STAT
• STAT+: The ESMO presentation everyone’s been waiting for — STAT
• Opinion: Christina Applegate puts a face on one of the most consequential challenges in… — STAT
• Opinion: STAT readers on Elizabeth Holmes, cancer screening, and more — STAT
• Healthcare AI needs a different model-development flywheel — HN-HC
• Inside The Medical AI Factory — HN-MED