Linger, a new open-source personal agent, runs on consumer PCs. It handles memory, agent teams and email triage locally using a MoE engine that fits a 125B model onto a 12 GB GPU.
A NeurIPS paper looks inside looped language models and finds that each pass through the shared block repeats the same processing stages a normal Transformer runs once. Whether those stages hold steady depends on a few design choices.
Monash and German researchers prune the Hydra and Quant time series classifiers, cutting inference energy by up to 70% with only a small loss in accuracy, and show that hardware choice can change which model looks best.
A revised MIT/Oxford paper shows that tuning AI classifiers for accuracy can quadruple worst-case liability, and proposes insurance contracts that force safer decision thresholds.
Meituan's open-weight talking-avatar model now holds a consistent face for more than 10 minutes, generates in 8 steps, and placed second only to OmniHuman-2.0 in Meituan's own human-preference vote.
Paweł Huryn's free, MIT-licensed product-engineering pack gives coding agents 45 skills that take a product from first idea to paying users, and no session is allowed to review its own code.
MI-Pruner decides which image tokens to drop by measuring how much each one shares with the text prompt. It never reads attention maps, so it can be added to off-the-shelf multimodal models without changing them.
TurinTech's Artemis GGUF build packs Google's Gemma 4 E2B into 2.05 to 2.31 GB, and it holds up much better than other public files in that size range.
ai-coustics' Quail Voice Focus 2.3 keeps the main speaker even when they're quiet or far from the mic. It runs on CPU in about 30 ms and cuts word errors by up to 43%.
Sellbird, an open-source AI sales agent that runs on a Claude or ChatGPT subscription you already pay for, is climbing GitHub's trending list with 6.4k stars.
In a 10-person exoskeleton pilot, walking patterns fitted to each user scored no better than an averaged pattern. How long people had been in the device mattered more.
Red Hat's 4-bit NVFP4 build of Gemma 4 26B-A4B cuts memory needs by about 75%. It keeps 97 to 99% of the original model's accuracy on most benchmarks, but agentic tool use and hard math take bigger hits.
Pooling more training corpora can make speech deepfake detectors worse. IDFE adds a corpus classifier with gradient reversal so the model can't lean on dataset quirks, and it cuts pooled error by up to 20%.
The ZazenCodes Season 3 repo is trending because of its Claude Code "software factory" templates. They turn labeled GitHub issues into reviewed pull requests, and a human still does the final merge.
An open-sourced agent skill methodically hill-climbs local LLM serving speed across any hardware and quantization, with measured gains up to 2.1x on prefill.
Frontier coding agents found 'significant' effects in data that contained none in up to 98% of runs. A new replay test corrects the p-values by measuring the agent's own selection bias.
A three-sensor neural network trained on simulation data reconstructs the full temperature and flow fields of a real natural circulation nuclear test rig.
A pixel-space diffusion transformer from CMU and Rice learns to merge related image patches by itself. It scores 1.48 FID on ImageNet and needs 2.4× less training compute than JiT to reach the same quality.
Nanonets-OCR2-3B is a compact vision-language model that turns messy documents into clean, LLM-ready markdown with semantic tags for tables, equations, signatures, and more.
A self-hosted, Apache-2.0 alternative to ChatGPT Enterprise that puts shared agents, RAG, a real browser and 4,700+ MCP servers behind your own firewall.
A multi-agent reinforcement learning system from DeepReinforce swept three consecutive live Codeforces rounds, outperforming every human including legendary grandmasters.
A new paper argues that self-attention works like the Dyson equation from many-body physics, which would explain why sum-pooling models break down on correlated quantum states.
AtomicChat shipped refusal-ablated GGUF builds of Liquid AI's LFM2.5-24B-A2B MoE, cutting English refusals from 96.3% to 1.2% while gaining MMLU points.
ENIGMA-360 pairs 180 headset videos with 180 synced fixed-camera videos of real repair work. Top video models break down when they have to switch between the two views.
A local-first Electron app translates English PDFs into side-by-side Chinese, pairs every paragraph with an AI chat, and plugs into Claude Code or Codex for annotations.
A community-trained LoRA for Krea 2 focused on Asian character generation has crossed one million downloads, pulling the open-weights base model into the spotlight.
A new Transformer layout runs full blocks side by side. In small MoE tests it beats standard stacks on quality at matched compute and decodes up to 51% faster.
ChatR Labs ships a rank-128 LoRA merged into Qwen-Image-2.1 across GGUF, FP8, INT8, NVFP4 and NF4 builds, with an uncensored tune that improves adult-prompt adherence while running on 4 to 14 GB of VRAM.
FedBNN trains one-bit neural networks across federated devices using rotation-aware binarization, cutting FLOPs up to 58x and memory 32x with competitive accuracy.
A trending open-source project turns Obsidian into a persistent memory layer for Claude Code, Codex, and Google Antigravity, with no API keys or servers required.
A new review paper shows that noise contrastive estimation, bridge sampling, and importance sampling are the same algorithm in disguise, and uses that insight to build better estimators for energy based models.