2026-07-21

An example TL;DR I have open-sourced NInfer , a from-scratch C++/CUDA inference engine currently specialized for two exact Qwen3.6 checkpoints on a single RTX 5090. Both the engine and the converted model artifacts are publicly available: Github : The main result: Qwen3.6-35B-A3B sustained 542 tok/s while generating a full 65,536 token completion, on a single RTX 5090, single request. My goal was to find out how fast inference can get on a single GPU (in my case RTX 5090), with a fixed model and

Two critical security flaws in WordPress’ software have given hackers the chance to remotely take over tens of millions of websites, according to an estimate by a cybersecurity researcher.

Under the new system, the protocol will take a looser, "stateless" approach to session IDs on the server side, similar to how most ordinary websites already work.

Edinburgh-based tech firm Craneware said customer data was stolen during a cyberattack. The company makes software that thousands of U.S. hospitals, pharmacies, and clinics rely on for billing patients, potentially exposing health data.

Hugging Face is urging users to rotate any access tokens stored on the platform and review account activity.

Governments look at banning ransom payments in face of increasingly sophisticated threats.

OpenAI shares lessons from deploying long-running AI models, highlighting new safety risks, observed failures, and improved safeguards through iterative deployment.

Augment Code's Vinay Perneti talks models, harnesses, and context.

Recent robot foundation models operate with single-step or short-history visuomotor context. We introduce Test-Time-Training Robot Policies (RoboTTT), a robot model and training recipe that scale visuomotor context to 8K timesteps, three orders of magnitude beyond state-of-the-art policies, without growing inference latency. At this context length, we unlock new robot capabilities: one-shot in-context imitation from human video demonstrations, on-the-fly policy improvement, robustness to perturb

Reinforcement learning (RL) has become central to improving large language models (LLMs) on complex reasoning tasks, yet RL post-training is largely studied in isolation from the pretraining that precedes it. As a result, two basic questions remain open: (1) how do pretraining choices (model size, data) shape the returns to RL compute, and (2) what does RL actually do to the model? These questions are difficult to study in the standard LLM setting: pretraining corpora are vast and uncontrolled,

Under model--harness co-evolution, harnesses are not merely inference-time scaffolds but data-generating components whose execution traces can shape future foundation models. This motivates harness-in-the-loop learning: optimizing harnesses for both immediate agent performance and the quality of traces used for future model training. However, continually updating provider-built scaffolds is costly and labor-intensive. We therefore investigate whether optimizing user-constructed harnesses in a ta

I asked myself where the Bonsai models actually land, so I ran them and compared to the results I already have for qwen-3.6-35b-a3b and qwen-3.5-9b on the same harness. Thought it might interest more people. Setup: little-coder harness via the harbor adapter, all 89 tasks of terminal-bench 2.0, single attempt (k=1), 40-turn cap, temp 0.2. RTX 5070 Laptop 8GB, i9-14900HX, 32GB RAM, CUDA 13.1. Runtime is PrismML's llama.cpp fork (stock llama.cpp can't load the 2-bit kernels). Results: Ternary-Bons

Hey guys, we collaborated with AMD to enable you to train, run, and deploy LLMs across nearly all AMD hardware including Radeon, Instinct, Ryzen, and data center GPUs. It works on Windows , WSL, and Linux and we have optimized ROCm builds for both training and inference. If you don't know about Unsloth , we're a fully open-source local UI that enables you to do pretty much anything with local models (RAG, chat, train, coding etc)! For those who don't have AMD GPUs and only CPUs, we still support

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Over the weekend, several current and former advisors to President Donald Trump on AI publicly lobbed insults at the country’s leading AI companies. David Sacks, the president’s AI and crypto “czar” until…

Large language models (LLMs) are transforming recommender systems from matching co-occurrence patterns in historical behavior toward reasoning about the intent that drives it. RecGPT-V1 pioneered this paradigm on Taobao by centering user understanding, and RecGPT-V2 scaled it via coordinated multi-agent reasoning; both are deployed in production with consistent gains in user experience and commercial outcomes. However, operating RecGPT at scale reveals three challenges: (1) stateless behavior mo

arXiv:2607.15367v1 Announce Type: new Abstract: Desktop voice assistants are still dominated by cloud pipelines that ship raw audio off the machine and expose a fixed set of skills. We describe AnovaX, a small local-first assistant that runs entirely on the user's computer and treats the desktop itself as its action surface. A single Python process wires together a wake-word gate, a speech pipeline, an LLM planner (Gemini) that emits a JSON plan of tool calls, a whitelist-and-denylist safety lay

arXiv:2607.15660v1 Announce Type: new Abstract: While LLM agents demonstrate strong reasoning abilities in compact and well-defined scenarios, they struggle to maintain robustness and effectiveness when faced with large-scale, diverse, and dynamic real-world environments that demand seamless tool integration. To address this gap, we introduce ToolVerse, a comprehensive framework that scales up agentic RL environments and enables agents to perform complex long-horizon reasoning in Tool-Integrated

arXiv:2607.15280v1 Announce Type: new Abstract: Sequential diagnosis requires balancing diagnostic accuracy against resource costs through iterative information gathering. Existing Large Language Model (LLM) approaches exhibit a critical knowledge-reasoning gap: despite encoding extensive medical knowledge, they struggle to reason systematically under cost constraints, often resorting to excessive testing. We propose GraphDx, a knowledge-enhanced framework with two core innovations. First, we de

Looped Transformers scale sequential computation by applying a compact stack of physical blocks for multiple rounds, increasing unrolled depth without increasing stored parameters. This reuse changes the residual-scaling problem: in an untied Transformer, each residual branch receives and applies its own parameter update, whereas in a looped Transformer one shared update aggregates gradients from repeated visits and is read back by those same visits in the next linearized forward pass. We formal

According to source, it is the locally ranked AI model, the best among 4b models Source :

Hello everyone, I wanted to share a recent project of mine, which brings a 13.1 million parameter convolution transformer model to a Thanks to quantization, this model now fits into 14mb of flash memory and it now sits at 256kb of SRAM as well as 4mb of PSRAM to transcribe 8 seconds of audio. The speed is still painfully slow. It is lightning fast compared to my initial attempt however, which took 10 minutes of inference time to transcribe 5 seconds of audio. I also gave the whisper tiny model a

Hyper-Connections (HC) expand the residual stream of Transformers into N parallel streams, providing a form of memory scaling beyond model width and depth. Manifold-Constrained HC (mHC) stabilizes this formulation at scale. The large gains from N{=}1 to N{=}4 suggest residual-stream expansion as a promising scaling axis. However, existing HC-family methods typically stop at N{=}4. Our experiments reveal why: scaling mHC beyond this point yields diminishing performance gains and rapidly increasin

I haven't kept up since around February so I'm just not even sure... and there are quite a few options. I don't care about parameter size, from tiny to huge, what matters most is performance, I just want all the best safely locally stored, I'll worry about running them later. So, what do you consider some of the best of the best currently? Whether highly specialized, giant do everything well, or anywhere in between Edit: Also what you use any specific models for or the best you've found for any

arXiv:2607.15550v1 Announce Type: new Abstract: Mobile graphical user interface (GUI) agents have demonstrated remarkable capabilities in automating complex tasks, yet they introduce critical safety risks where a single erroneous action can lead to irreversible consequences. Existing safety mechanisms are primarily reactive, lacking the ability to assess risks before execution. In this paper, we introduce SeerGuard, a consequence-aware safety framework designed to mitigate these risks through pr

Open weights 975B multimodal model built for fine-tuning Discussion | Link

So, im 50 and autistic. I learned C++ in college when I was young and didnt really ever use it that much. I also learned a bit of assembly from using cheat engine for game hacking (dont worry! single player or private servers!) which meant I got a little LUA experience. I've always disliked coding. Its such a "bang head against wall" type activity but I got into it because its pretty cool being able to make a computer do what you want. This entire attitude died a LONG time ago because coding was

​I'm looking to run a very lightweight local model that acts as the brain, handling the logic and comprehension, while hooking it up to a web search tool to act as its memory and knowledge base.

arXiv:2607.15459v1 Announce Type: new Abstract: A trained deep reinforcement learning policy is a black box, and we ask whether it can be made explainable by rewriting it as an executable logic program that reproduces its behaviour and that a person can read, a logic engine can run, and an optimizer can edit. We present a three-stage post-hoc transformation that extracts a frozen proximal policy optimization teacher, induces an ordered rule list from its decisions in the manner of classical rela

Article URL: Comments URL: Points: 30 # Comments: 22

It may seem counterintuitive, but 113-degree water is still cool enough to cool NVIDA's latest hardware.

So I've been working on this app for a very long time now. Started off by wanting an app that separates long form notes from short form notes visually and keeps everything organized on a canvas. The app features: -Sticky notes for quick, small notes and A4 notes for long form documents -A quick mode that uses local storage to quickly open the app and jot down a thought on the canvas -A PDF export option that exports your notes into 3 custom designed PDF styles -Visual hierarchy to replicate file

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. AI is more likely than humans to form biases when hiring The next time you apply for a job, AI may screen your résumé before any human sees it. But there’s…

We have been having extensive discussions around open source strategy. We will discuss it more at our next board meeting, but one thing we’d like to do soon is to create a language model with the approximate capability of GPT-3 that can run locally on consumer hardware and release that. We’d like to do it soon, before Stability or someone else does. In general, we think this helps discourage others from releasing similarly-powerful models, and makes it harder for new efforts to get funded. — Sam

Your personal context file for every agent. Discussion | Link

the open ai exec in his "ai communism" post suggested a fraudulent FUD campaign and trump executice order against open source / chinese models. unfortunately that seems pretty likely to happen. i have always use huggingface for downloading models and they would be forced to comply. are there any established alternatives or mirrors that aren't us-based?