./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.

Saturday, 2026-09-12·last updated 2026-09-12 23:56 UTC·9/9 feeds ok

[ hardware ]12

• b10933 — llama.cpp
• b10932 — llama.cpp
• d-Matrix Joins the NVIDIA NVLink Fusion Platform — STH
• b10931 — llama.cpp
• b10930 — llama.cpp
• b10929 — llama.cpp
• b10927 — llama.cpp
• b10926 — llama.cpp
• b10924 — llama.cpp
• b10923 — llama.cpp
• b10922 — llama.cpp
• Qwen3.8 Flash Next llama.cpp config tuning — r/LL

[ new models ]8

• For those of you forced to only use open models from Western labs in production, what ar… — r/LL
• Anybody use frontier models like Astra/Fable for planning/judging, and qwen3.8 as the ma… — r/LL
• 3.8-27B has ruined 3.5/3.6-35B’s for me. It’s just *absurdly* superior. — r/LL
• Agnes-AI/Agnes-3.0-Flash 33B Multimodal, AA score: 36 — r/LL
• Concerning "humanlike models" and chatbot RP in general... — r/LL
• Unsloth UD-quants - Qwen 3.8 27b for example - worth using 8-bit or stick with faster 6… — r/LL
• Qwen-Next seems worse to me then 3.8 27b for coding, but I feel like I must be missing s… — r/LL
• Rebuilding AUTOMATIC1111 with Gradio Workflow — HF

[ local ai ]12

• This seems more probable than it was before. — r/LL
• Qwen3.8 Flash Next now at 1.2k t/s prefill on Strix Halo — r/LL
• Real-SWE Benchmark (new) — r/LL
• M2 Ultra/Qwen3.8 Flash Next Update - latest oMLX introduces substantial speedup — r/LL
• Looks like a coordination to stop distribution of intelligence — r/LL
• This draft model is OP on 16 GB cards for Qwen 3.8 27b — r/LL
• Intel Linux NPU driver only now officially supports Ubuntu 26.04 LTS — r/LL
• bartowski/Qwen3.8-27B-GGUF · Hugging Face - Updated (Per-tensor layout) — r/LL
• I am impressed and I owe you one, Qwen 3.8 flash next (vision)! — r/LL
• ggml-cuda: hip: add missing AMD GCN MMQ config by thelittlefireman · Pull Request #27841… — r/LL
• Antirez Deepseek 4.1 flash gguf on HF — r/LL
• I fine-tuned a 2B LLM on our WhatsApp group chat, and shared how to do it on GitHub as a… — r/LL

[ benchmarks & papers ]2

• Releasing smolbenchmark: Helps you choose the best model for your hardware! — r/LL
• A collaborative agent with two lightweight synergistic models for autonomous crystal mat… — NatMI

[ tools ]2

• tencent/AuK-Flash · Hugging Face — r/LL
• Training a 3.8B LLM to 0.384 CORE for $998 — HN

[ healthcare ai ]10

• STAT+: Breast cancer pill from AstraZeneca misses mark in pivotal trial — STAT
• STAT+: Scholar Rock wins FDA approval for first drug to target SMA muscle loss — STAT
• STAT+: Pharma wins a battle over the cost of wastewater treatment in Europe — STAT
• U.S. officials declare an end to record-setting cyclospora outbreak, but its origin is s… — STAT
• STAT+: Up and down the ladder: The latest comings and goings — STAT
• Medicaid will let states use ‘tiers’ to determine medical frailty — STAT
• STAT+: Pharmalittle: We’re reading about anxious Novartis investors, patent expirations… — STAT
• STAT+: Cancer drug shortages keep disrupting patient care — STAT
• Healthcare AI's next test is integration — HN-HC
• Inside The Medical AI Factory — HN-MED