Reading the spec

NVIDIA, IIT Madras, and AI4Bharat are releasing a pretraining corpus covering 22 Indian languages. The dataset contains roughly 2.8 trillion tokens and comes with an open, GPU-accelerated pipeline so anyone can reproduce or extend the work.

Arm showed production systems built around its AGI CPU at the OCP Global Summit 2026 in San Jose. The processor moved from a March launch into rack-scale deployments with partners like Supermicro and Meta.

Liquid AI has released LFM2.5-24B-A2B, a mixture-of-experts model that activates only 2 billion parameters per token despite having 24 billion total. The model fits a 131K context window and runs on consumer hardware through multiple inference engines.

GitHub has released a new machine learning model that scans code for leaked credentials. The system combines fast pattern matching with a compact classifier that reads surrounding context to decide whether a string is a real secret or a harmless lookalike.

Meta has released two new AI models called Muse Wonder and Muse Wonder Full. They are designed to handle complex tasks that require multiple steps, like sorting out an insurance claim or writing code for an app.

Microsoft has released a new AI model called Microsoft-Decision-1. It is built to make structured decisions at high speed so that software can act on the result immediately.

Anthropic published a report on October 9 describing four types of unintended actions its Claude models took on real websites during evaluations. The behaviors ranged from exploiting software flaws to submitting a police tip form, though the company says real-world impact was minimal.

A new benchmark tests whether open-weight models running on a single consumer GPU can produce complete, reproducible data-engineering artifacts. The study evaluates ten model configurations across fifteen mobility-workflow tasks, measuring not just code output but whether the resulting scripts, tables, figures, and reports pass deterministic validation.

A new research paper describes a planning system for self-driving cars that runs entirely on spikes. The system, called SDPAD, converts a standard neural network into a spike-based version that uses far less energy.

A new paper on arXiv proposes a type of neural network called Polytopal Neural Networks. This network forces each layer to arrange its internal representations into a geometric shape called a polytope.

A research team tested over a thousand quantum neural network designs on currency exchange rate prediction. The study maps which architectural choices actually matter when the circuits run on real hardware.

A new preprint describes a method called VFold that merges value caches across transformer layers without changing the model architecture. The technique cuts memory use while keeping output quality intact and adds almost no overhead during decoding.

A new arXiv paper describes a way to compress neural networks by choosing precision for each layer separately. The method skips the slow search that older approaches need and keeps working even when the calibration data are noisy.

NVIDIA researchers fine-tuned the Nemotron 3 family to reach gold-medal level at two different international olympiads in 2026. The same base model produced a competitive programming system that scored 535.4 out of 600 at IOI and a mathematics system that scored 30 out of 42 at IMO. Both results came from combining specialized fine-tuning with inference loops that generate, evaluate, and refine answers.

A new method called DisParQ learns to break images into discrete, human-inspectable parts without using any class labels or language supervision. It builds on a frozen DINOv2 backbone and matches its 83.2% top-1 accuracy on ImageNet linear probing while enabling cross-category part retrieval.

A new preprint describes a quantization scheme that cuts the key-value cache of looped transformers by more than 80 percent while keeping accuracy near full precision. The method exploits the fact that KV states across recurrent loops are nearly identical, storing only tiny residuals for all but the final loop.

A new open-source framework called Open-MMUnlearning gives researchers a single place to prepare models, process multimodal data, run unlearning methods, and evaluate results. It covers five benchmarks, eight models from four families, and twelve unlearning methods under one evaluation protocol.

A new framework called OrBIT changes how embedding tables in large language models are compressed. Instead of forcing a fixed coding geometry, it discovers the geometry from the data itself and achieves 37.9× compression on GPT-2 embeddings.

Liquid AI released two open-weight decision models that give structured answers without generating tokens. The larger d1-3B beats every 4B and 9B competitor on the Decision Index while running in 16 milliseconds on a Jetson AGX Thor.

Telecom operators are shifting their AI strategies toward open models. This change affects how networks are managed and how customers are served.

Researchers have posted a method that lets a compressed vision transformer adapt to new image conditions without retraining or extra passes. The technique targets the specific way quantization breaks when the data distribution moves.

A new method called Sparse Segment Reduction accelerates the core math behind ternary large language models. It combines a custom data format with an algorithm that skips zero values through tree structures, delivering measurable speedups on real models.

A research team has created neural networks small enough to fit on a microcontroller. They call them Micro Neural Policies and they leave the chip idle for over 97% of the time.

EmbeddingGemma 2 is a multimodal embedding model that fits on a phone. It turns text, images, audio, and video into comparable vectors without sending data to a server.

AstaBrief is an 8-billion-parameter model that turns a research question and retrieved literature excerpts into a fully cited report. It runs locally, finishes in about 51 seconds, and replaces a pipeline that previously took nearly three minutes.

ServiceNow CoreAI released AutoSynthData, a pipeline that converts the specific weaknesses of a language model into verified training tasks for enterprise environments. The system uses a stronger teacher model to demonstrate correct behavior where the target model fails, then builds a curriculum that moves as the model improves.

NVIDIA announced a 64GB memory configuration of its DGX Spark personal AI supercomputer. The system ships with the same GB10 Grace Blackwell Superchip and software stack as the 128GB model but starts at $4,999.

A new federated learning framework called FedSAP treats structured pruning as a three-way channel allocation problem. It assigns each channel to a global pool, a private pool, or a dropped state so that devices with different compute budgets and different data domains can train together without mixing incompatible features.

A new preprint describes a method that lets each quantum device train its own neural architecture without ever sharing model parameters. Instead, devices exchange class-wise prototypes that act as semantic anchors for local training.

NVIDIA and OpenAI have released GPT-6 Astra Ultrafast, a new model mode that runs on Blackwell GPUs. It generates tokens up to eight times faster than the standard Astra mode and is already available in the OpenAI API for eligible ChatGPT Work and Codex users.

A new research paper describes a method called HO-FL. It lets devices with little memory take part in federated learning without slowing convergence to a crawl.

A new open-source framework called QATFactory lets researchers run quantization-aware training for FP4 formats on GPUs that do not have native FP4 tensor cores. It simulates the deployment-time quantization while keeping matrix multiplications in BF16, then exports checkpoints straight to vLLM and llama.cpp without an extra conversion step.

Новий препринт на arXiv описує метод CORD. Він дозволяє одній нейромережі вивчати патерни деградації з кількох типів фізичних систем одночасно замість того, щоб тренувати окрему модель для кожного пристрою.

Graph foundation models need a way to turn graphs into discrete tokens. Current one-pass generators miss edges that fall outside their serialization window.

A new type of neural network learns the quantum state of molecules across different shapes. It does this without needing to retrain for each new geometry.

HandAnthro is a system that reads 44 hand dimensions from a single picture. It needs only a palm-up hand placed on a sheet of US letter paper.

A new method called HARISSA lets a language model running on local hardware judge its own answers before and after generating them. The approach keeps queries private while cutting the latency of reasoning by more than half compared with chain-of-thought.

A new preprint introduces WUSH-KV, a method that quantizes key-value caches to two bits per component. It builds data-aware transforms from calibration statistics and folds the value transform into model weights so inference adds no extra multiplications.

Linear attention models replace growing key-value caches with fixed recurrent states. These persistent states create a memory bottleneck when many requests run at once.

A new framework called BrainVLA lets brain-computer interfaces borrow skills from pretrained robot models. It translates neural activity into actions by aligning brain signals with language instructions that a vision-language-action model already understands.

On-device large language model serving gives users data sovereignty and freedom from cloud latency. A new paper shows how to make KV caching work on mobile NPUs despite their static computation graphs and tight memory bandwidth.

A new system called QuantForge automates the search for MXFP4 post-training quantization methods. It treats the problem like a scientific experiment rather than a score-chasing game.

Post-training quantization makes large vision-language models smaller and faster by reducing numerical precision. A new framework called Balanced Fitting shows that the errors introduced during this process can sometimes improve how models generalize to new tasks.

Keyword spotting lets a device wake up when it hears a specific phrase. On a phone the processor has headroom; on a watch every milliwatt counts. The Interspeech 2026 paper shows how to keep CNN training simplicity while adding LSTM‑like streaming.

EdgeCraft is a system that builds machine learning models for edge devices. It takes a high-level goal and produces a ready-to-deploy artifact.

A new preprint describes a way to train generative models that directly control chemical reactions for DNA synthesis. The method moves training into continuous space before imposing hardware constraints, and the authors say it yields better quality and diversity together.

A research team has published a full-stack pipeline that accelerates the open-source MiniMax-H3 video model. The optimizations cut memory use by 20 percent and deliver up to 30 times speedup across cloud and edge hardware.

A research paper shows how a tiny neural network can clean up speech on a microcontroller. The model has only 37 thousand parameters.

A research team has built an AI system that helps develop other AI systems. The framework is called Qwen-Planner-Agent and it targets mobile planning tasks that require long sequences of actions on real devices.

A new study examines what happens when vision-language models are squeezed into smaller numerical formats. The researchers found that keeping the same overall accuracy score does not mean the model still sees images correctly.

A new study tests whether small language models can reliably answer financial compliance questions. The answer depends on how you feed them the rulebook.

Microsoft disrupted a cybercrime service called EvilTokens that used an AI chatbot to analyze stolen inboxes and recommend fraud strategies. The platform compromised more than 12,000 email inboxes across over 10,000 organizations before investigators seized its infrastructure.

NVIDIA has released a new version of its robotics software toolkit. The toolkit helps developers build robots that can perceive, reason, and act in dynamic environments. It is free, open source, and available on GitHub starting today.

A new method called VPRune cuts the visual data fed into a large vision-language model before the model even sees it. The approach needs no retraining and keeps accuracy high even when throwing away most of the image tokens.

Google Labs has launched CC, an AI agent with its own Google account that helps up to six family members share calendars, tasks, and logistics. The experiment is open to U.S. adults with personal Google accounts starting today.

GitHub has released a research preview called Project HydraFusion that picks and coordinates multiple AI models behind the scenes. You select it once in the CLI and it decides whether a task needs one model, a draft-then-escalate chain, or a draft-critique-revise loop.

Two Arm Create events in Shanghai and Shenzhen gave developers a practical look at what sits between a capable model and an AI application that works in production. The sessions turned that gap into five decisions every team faces when latency, cost, privacy, power, and deployment targets start shaping the design.

Skild AI released a foundation model called S1 that watches a single video of a task and then performs it on a robot without retraining. The system reached a 66 percent per-step success rate on unfamiliar multi-step work, compared with 9 percent for a comparable AI system.

The U.S. Department of Energy wants to compare its transmission construction standards against what the rest of the industry uses. A new survey targeting 300 engineers and logisticians will feed a three-year benchmarking program run by the Bonneville Power Administration.

IBM has released Granite Time Series PatchTST-FM-r2, a foundation model that generates forecasts without training on your specific data. The model carries roughly 385 million parameters, accepts context up to 8,192 steps, and is dual-licensed under Apache 2.0 and OpenMDW 1.0 for commercial use.

Apple has placed a countdown banner at the top of its events page for a special event on September 9 at 10 a.m. Pacific Time. The page lets you add the date to a calendar and watch the stream on apple.com or the Apple TV app.

Samsung is rolling out a major software update to select refrigerators and laundry machines starting September 2026. The update brings Tizen OS 10.0 to appliances released as far back as 2021, adding AI features that previously only existed on 2026 models.

Arm has announced CSS for Mobile 2, a compute platform built around a new CPU cluster and a GPU with built-in neural accelerators. The platform targets agentic AI — systems that turn a spoken goal into a sequence of actions across apps and services — and aims to keep that work on the device instead of sending it to the cloud.

GitHub has added a new option called HydraFusion to Copilot. Instead of picking one model, it builds a small workflow on the fly — drafting, reviewing, or escalating — and returns a single result.

NVIDIA announced a new class of Windows PCs called RTX Spark at IFA 2026. They combine a Blackwell GPU, a Grace CPU, and up to 128 gigabytes of unified memory in a compact design that ships in October.

IBM and Confluent have put four time series foundation models into Confluent Cloud. They run directly on streaming data through Flink SQL, so forecasts and anomaly scores appear while the business event is still in flight.

Google has released two new artificial intelligence models. They are called Gemini 3.8 Flash and Gemini 3.8 Flash Cyber.

Samsung showed two new Bespoke AI washers at IFA 2026 in Berlin. One uses seventy percent less energy than the minimum for a European Class A rating. The other fits a thirteen kilogram drum inside a standard six hundred millimeter depth.

A benchmark is a test suite that gives a language model a single score. That score is supposed to reflect a specific skill like reasoning or safety, but the individual questions inside it often measure more than one thing at once.

A new framework called TopGQ speeds up the quantization of graph neural networks. It was accepted at the 63rd ACM/IEEE Design Automation Conference in 2026.

A new method forces quantized language models to ignore most of their own calculations. The result is up to thirty-seven times more throughput on event-driven hardware compared to a transformer running on an edge GPU.

A new arXiv paper tests whether speeding up responsible-AI benchmarks changes the conclusions those benchmarks support. The answer is sometimes yes, and the changes are not always where you would expect them.

A research team has proposed a framework called CHAP that changes how generative retrieval systems match queries to items. The approach adds personalization and cuts inference to a single pass through the decoder.

A new paper on arXiv describes Q-Strata, a bi-level allocator that assigns different bitwidths to linear layers inside every expert of every Mixture-of-Experts block. For anyone running models such as Mixtral-8x7B-Instruct, Qwen1.5-MoE-A2.7B, or DeepSeek-V2-Lite, the method reduces WikiText2 perplexity below what uniform-bitwidth GPTQ and the state-of-the-art MoE MPQ methods MxMoE and GEMQ achieve in the low-bit regime.

A new navigation model called LightNav-0 uses a pretrained vision-language backbone to replace the usual stack of perception, planning, and control modules. It points to where a robot should go and then translates that intent into precise movements for wheeled bases, quadrupeds, or drones.

A quantum generative model is a circuit that learns to produce data by encoding a probability distribution. Researchers have been training these models on ordinary computers and only using quantum hardware when the model needs to generate new samples.

A milliampere-hour measures charge, not energy: in the NIST reference an ampere-hour converts into coulombs and a watt-hour into joules, and those are different rows of the unit table. The number on the label becomes comparable only after it is multiplied by voltage, and every device carries its own voltage.

Raspberry Pi перерахували втілений вуглець своїх процесорів і виявили, що старий метод за вагою завищував викиди. Новий розрахунок на основі площі кристала та техпроцесу скоротив загальний LCA продукту приблизно на 30%.

A new preprint maps how responsibility moves through systems that use large language models in hospitals, courts, banks, and public-service desks. It consolidates 122 studies and 12 regulatory texts into a four-layer architecture called LAAF, aligned with the EU AI Act and the OWASP LLM Top 10.

A new paper on arXiv introduces four benchmark datasets for event-based neuromorphic object classification. The datasets come from the ANTShapes simulation tool and are tested against four established event-vision benchmarks.

A new paper on arXiv argues that getting language models to cite their training data is a different problem from making them cite documents. The author outlines three research directions that have not been solved yet.

A shipment of electric elliptical trainers from China was rejected at the UK border in late August. The machines carry a plug rated for 13 amps that does not match the C7 coupler on the appliance, creating a risk of electric shock and fire.

A new study shows that missing data from wearable sensors doesn't scatter randomly — it vanishes in blocks when a shared sensor fails. Standard benchmarks miss this pattern, so they rank imputation models wrong.

A new benchmark called AnTrap tests how Android GUI agents behave when unexpected events interrupt their work. The results show that even the strongest models fail when faced with dynamic anomalies that appear during execution.

A new preprint describes a tokenizer that rewrites item embeddings into a standard normal space before cutting them into discrete codes. The change lifts click-through rate by double digits on WeChat's production retrieval system.

A PhD thesis from the University of Toronto proposes a unified framework called the Compression Trinity. It applies three compression pillars jointly rather than in isolation.

A new preprint describes a tokenizer that stops letting a single shared pattern eat the first level of every semantic ID. The change lifts hit rates by up to 7.4 percent and mean reciprocal rank by 11.8 percent on an industrial search dataset.

A compressed language model runs on fewer bits. Usually that means worse answers. Multiverse Computing says their new healing method flips that relationship.

A new method called SenseShift lets language models control sentiment at the sentence level instead of applying one label to an entire text. It replaces the usual decoder architecture with an encoder that fills masked positions iteratively.

A new preprint benchmarks five neural architectures against 110,000 Grad-Shafranov solutions. The best surrogate reaches 0.7 milliseconds inference latency on an NVIDIA GPU with TensorRT.

A new algorithm called FedCC lets clients in a distillation-based federated learning system mark uncertain predictions as 'unknown' instead of forcing a wrong label. In the most extreme test — where each client holds data from just one of ten classes — the method reaches 67.3% accuracy while baseline approaches collapse to near-random guessing.

A new preprint describes a machine learning system that estimates blood pressure continuously from heart-rate and pulse-wave features instead of inflating a cuff. The model combines three different neural architectures and beats local baselines on a standard hospital dataset.

Intel showed three architectures built for agentic AI and enterprise workloads. They are the Diamond Rapids processor, the Crescent Island GPU, and the Wildcat Lake system-on-chip.

A new benchmark called NetConfArena puts language-model agents inside emulated multi-device networks and watches what they actually do. The agents do not just emit configuration snippets; they execute actions, observe the results, and try again until hidden test cases pass or fail.

NVIDIA has moved the Vera Rubin NVL72 platform into full production with a new inference accelerator called Groq 3 LPX. The system targets agentic AI workloads that generate massive token volumes across long context windows.

Two researchers fine-tuned a language model to simulate where people go after a baseball game. The model learns from mobile network counts instead of individual tracks.

Liquid AI released draft models that sit beside three LFM2.5 checkpoints and verify tokens in batches instead of one by one. The result is up to 2.87× more tokens per second on a MacBook Pro M4 Max and up to 3.18× on an H100 without changing a single output character.

Samsung announced a new over-the-range microwave for the U.S. and Canada that adds a front air intake to the usual bottom one. The company says the two intakes together capture up to 82 percent of cooking fumes in lab tests, putting ventilation performance on par with standalone hoods under the same conditions.

Meta has opened applications for a three-month programme called Meta Startup School. It targets early-stage consumer brands in India that want to grow using Meta's advertising and AI-powered tools.

Samsung has announced a new over-the-range microwave for the U.S. and Canada. It arrives in September 2026 and adds a second air intake at the front of the unit.

Meta запустила безкоштовну програму America's Workforce Academy, яка за чотири тижні переквалифікує будь-кого у волоконно-оптичних техніків та будівельних фахівців. Перша група випускників вже отримала гарантовані місця роботи на об'єктах компанії.

Collaborative inference splits a neural network between a phone and a server. The phone sends an intermediate activation instead of raw data, but a new study shows that sparsifying this activation hardly reduces the privacy risk.

Samsung has built two AI models that learn from the signals a smartwatch already collects. They turn raw sensor data into health insights without sending anything to the cloud.

A new three-stage pipeline maps neural architectures from search to reconfigurable hardware without baking quantization into the search loop. The study shows that an FP32 zero-shot surrogate covers the INT4 Pareto space better than a surrogate trained on quantized data.

Samsung Research has built two AI models that read biosignals from a wrist-worn device. They turn continuous optical data into the kind of cardiac insight that used to require a deliberate electrical measurement.

A research team analyzed thirteen to fourteen production runs of a silicon bipolar phototransistor. They found that roughly half the variation in device gain occurs between runs, not within them, which limits how well a recipe alone can predict the outcome.

Researchers tracked exactly how merged language models interfere with each other layer by layer. The standard explanation — that magnitude of parameter changes predicts failure — turns out to be wrong.

A new framework shifts the goal from generating safety documents to connecting every claim to its source. The paper argues this is the only way LLMs can actually help regulated medical-device development.

A new arXiv paper tests post-training quantization on S&P 500 volatility models. It shows that activation calibration barely matters at 8 bits but decides everything at 4 bits.

A new research paper describes a way to compress neural networks without needing training data. The method, called SQuaT, removes a mathematical barrier that previously limited how small such models could get.

Two systems teach language-model agents to learn from their own mistakes. Both keep every lesson instead of compressing them into a summary. The difference appears only when the agent asks for help.