RealFeedApp betaTop science & technology news, November 14, 2026
- Qwen3.8-27B runs locally on 17GB RAM/VRAM
The Dynamic V3.1 update enables running the 27B parameter model on 17GB setups with improved performance on Terminal Bench and DeepSWE benchmarks. Features include developer role support and MTP for faster inference.
- llama.cpp 1.0 adds FP8 context cache and safer llama-server
Speculative decoding accelerated token generation by 14–22% when handling multiple concurrent requests, while the 27B model achieved a 55% speedup. The update introduces API changes and adds FP8 context cache support.
- Two DGX Stations serve GLM-5.3 NVFP4 via ConnectX-8 RDMA
The 753-billion parameter GLM-5.3 NVFP4 model runs on two DGX Stations via vLLM, using tensor parallelism to keep all weights in GPU memory.
- Project Zero reduces vulnerability remediation timeline to 90 days
Google Project Zero reduced the disclosure window to 90 days because AI tools speed up vulnerability discovery. Researchers noted that the previous timeline is now insufficient for remediation.
- $300M Series C for LangChain
LangChain raised $300 million in a Series C round, doubling its valuation from $1.25 billion. Originally an open-source project for LLMs in 2022, the company now focuses on managing AI agents.
- Human Connectome data analyzed
Researchers analyzed Human Connectome Project data (N = 994) and found that spatial-temporal organization of brain networks reflects a state-trait continuum, linking stable personality traits to more stable dynamic patterns.
- VREfm ST80 strains distributed into two major clusters from 2016–2020
Analysis of 334 isolates revealed that the dominant lineage from 2012–2015 declined while other clusters rose between 2016 and 2020.
- Researchers improve genetic test for glaucoma
Researchers improved a polygenic risk score test using data from 6 million people to detect high-risk individuals for open-angle glaucoma.
- Transformers library loads ArsalanZ/Nenya-2.5B-A0.5B-Base model
The Nenya-2.5B-A0.5B base model, trained on 60 billion tokens, achieves performance comparable to Qwen3.5-2B trained on over 30 trillion tokens.
- Nvidia loses its biggest champion for AI startups
Jensen Huang's resignation removed a major investor from the startup ecosystem, leaving 105 rounds unfunded. This change could significantly impact dealmaking, valuations, and competition in the AI sector.
- SpaceX stops flying Falcon 9
SpaceX plans a major Starship test that could lead to discontinuing Falcon 9 flights, currently the sole provider of NASA astronaut transport to the ISS.
- Brain scans predict future memory problems
Stanford researchers found that functional MRI patterns in healthy older adults differ from those with Alzheimer's, enabling early prediction of memory loss.
- Sankaet Pathak bought a $4.6 million home in Sausalito
Foundation Future Industries missed payroll payments in October while its CEO borrowed $4.6 million to purchase a luxury property.
- Modeling shows how to read a qubit without measuring it
Researchers from China and Brazil developed a theoretical model to measure the electrical current produced by qubits without collapsing their superposition.
- Four Chinese AI labs use NVIDIA Blackwell chips barred from export
Forensic analysis of model weights reveals that DeepSeek, Moonshot, MiniMax, and Ant Group used NVIDIA Blackwell chips despite export bans. The researcher tested over six trillion combinations of chip layout and seed number to identify the hardware fingerprints.
- Mean contig length reached 18
Researchers applied PacBio HiFi long-read sequencing to reconstruct the genome of a P. falciparum isolate from Mali, revealing 301 pseudogenes and 18 contigs.
- Fewer bits mean more quantization errors
The study demonstrates that training large language models with 4-bit precision reduces memory usage and increases GPU speed, while introducing specific techniques to mitigate the impact of accumulated quantization errors on model accuracy.
- Beta Pictoris is roughly 20 million years old
The James Webb Space Telescope detected water and carbon dioxide ice in the disk surrounding the young star Beta Pictoris, which is approximately 20 million years old. The ice composition resembles that of comets in our own Solar System.
- RTX 3060 has 12GB memory for 1080p
The RTX 3060 remains the most popular GPU on Steam thanks to its 12GB memory and affordability, despite two generations of successors. In October 2025, it was used in 3.78% of surveyed PCs, tying with the RTX 5070.
- Frog-shaped robot swam 80 millimeters in 1 minute 54 seconds
Researchers from Denmark and Luxembourg built a gram-scale frog robot that moves using flashes of near-infrared light activating liquid crystal muscles on its hind legs. The robot completed a lap of 80 millimeters in an average of 1 minute 54 seconds.
