AI model news
Stories 361 to 419 of 419, from 6 September 2026 to 18 September 2026, newest first. The latest are on the first page →
-
Google launches Gemini 3.8 Live voice model with $0.023 per minute pricing
Google announced Gemini 3.8 Live on September 15, 2026, a voice AI model that enables continuous, natural conversation in 97 languages and can run background tasks such as search and tool invocation. The standard…
1 sourcenote.com -
inclusionAI releases Realtime-Venus 9B audio-visual interaction model
inclusionAI has published the Realtime-Venus system on Hugging Face, offering two checkpoints: Realtime-Venus-Omni, a 9 billion‑parameter audio‑visual model, and Realtime-Venus-Audio, an audio‑only variant. Both are…
1 source primary sourcehuggingface.co -
OpenAI launches Astra for Law, a GPT-6 legal AI platform with 230 million‑URL search index
OpenAI announced Astra for Law, a legal‑focused configuration of its GPT‑6 Astra model that pairs the model with a searchable index of more than 230 million U.S. case‑law, statute and regulation URLs, sourced in part…
7 sourcesThe DecoderLatent Spacethenews.com.pkzeebiz.com +3 more -
Speculative decoding explained: how AI speeds up text generation
Speculative decoding speeds up autoregressive generation by letting a lightweight draft model propose a block of tokens that a larger target model verifies in parallel, without altering the target distribution. The…
1 sourceUnite.AI -
Modality Discrepancy Transformer reaches 0.7408 Macro F1 on BAH dataset
A new paper on arXiv proposes the Modality Discrepancy Transformer (MDT) to detect ambivalence and hesitancy in clinical videos, where facial, vocal, and linguistic cues conflict. MDT expands a prior 6‑token design to…
1 source primary sourcearXiv cs.CL -
OpenAI’s GPT‑6 Astra 99.9% ARC‑AGI‑3 score may reflect benchmark harness, not AGI
OpenAI announced that its GPT‑6 Astra achieved a 99.9% score on the ARC‑AGI‑3 benchmark, a figure that quickly became a headline for the company’s push toward an AGI narrative. The nonprofit ARC Prize, which created…
1 sourcesrutiosocial.com -
IFM releases K2-Horizon-7B-Uno diffusion-augmented LLM, claims 5200 tps speedup
The Institute of Foundation Models (IFM) has posted a new 7‑billion‑parameter language model, K2‑Horizon‑7B‑Uno, on Hugging Face. The model adds a diffusion pathway to the existing K2‑Horizon‑7B architecture via…
1 source primary sourcehuggingface.co -
OpenAI's GPT-6 Astra: Changing Work and Research
OpenAI has unveiled GPT-6 Astra, its most advanced AI model, designed to revolutionize various industries. This model is state-of-the-art in areas such as computer use, software engineering, cybersecurity, and…
22 sources primary source HN 512Latent SpaceMarketing AI InstituteTom's Hardwarecryptobriefing.com +13 more -
DeepSeek releases V4.1-Flash with 1M context and MIT license
DeepSeek AI has released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model designed to address the memory bottlenecks of long-horizon agent workloads. The model features a 552B parameter backbone with 196B…
6 sources primary sourceKDnuggetsgithub.comUnite.AIfinance.yahoo.com +2 more -
FreedomIntelligence releases HuatuoGPT-3-9B medical LLM with open usage guides
FreedomIntelligence has made its new HuatuoGPT-3-9B model publicly available on Hugging Face. The 9‑billion‑parameter model is built on Qwen3.5‑9B and uses a One‑stage Policy Optimization (OnePO) technique that adapts…
1 source primary sourcehuggingface.co -
Google moves Chrome to a 2‑week update cycle to narrow patch gap
Google has shortened the release cadence for its Chrome browser from four weeks to two weeks, a change announced in a blog post and confirmed by Chrome lead Ben Mason. The move is aimed at shrinking the "N‑day" patch…
1 sourcezdnet.com -
Google launches Gemini 3.8 Live and Extended Thinking voice models
Google announced on September 15, 2026 the release of two speech‑to‑speech models—Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking—designed for real‑time voice agents. The models support near‑real‑time visual…
7 sources primary source HN 54Ben's BitesSocial Media ExaminerSimon WillisonMarkTechPost +3 more -
Salesforce launches Koa, its first CRM reasoning model built on NVIDIA Nemotron 3 Super
Salesforce has officially unveiled Koa, its first proprietary CRM reasoning model, developed in partnership with NVIDIA. The model is built by post-training NVIDIA’s Nemotron 3 Super on a synthetic dataset derived from…
6 sources primary sourceCNBC TechnologyNVIDIA Blogsiliconangle.comMarTech +2 more -
AI Labs Call for Pause Amid Recursive Self-Improvement Concerns
The leaders of major AI labs have called for a pause in technology development, warning that AI could soon improve on its own and slip beyond human control. This follows reports of swarms of AI agents breaching…
2 sources HN 26aol.comWired AI -
Apple launches iOS 27 with context-aware Siri AI overhaul
Apple has begun distributing iOS 27, marking the official arrival of its significantly upgraded Siri AI assistant. The update, released on September 14 across iPhone, iPad, Apple Watch, and Vision Pro, introduces a…
15 sources HN 25aol.comhothardware.comThe Decoderextremetech.com +11 more -
OpenRouter Token Usage Soars 25,000% Since Jan 2025
OpenRouter, a platform that lets developers plug AI models into their products, has seen an explosive rise in token consumption. Weekly usage jumped from 0.5 trillion tokens in January 2025 to 126.2 trillion tokens, a…
1 sourceThe Decoder -
Z.ai Deploys GLM‑5.3‑Flash on 100,000‑Chip Chinese Cluster
Z.ai announced a technical account detailing how it built a production‑grade inference service for its GLM‑5.3‑Flash model on a cluster of more than 100,000 Chinese‑made AI accelerators, a first at this scale. The…
1 sourceUnite.AI -
AI Chatbots Lag Behind on Recent Information
When testing popular AI chatbots like ChatGPT, Claude, and Gemini, I found that they often lack up-to-date knowledge about recent events. For instance, both ChatGPT and Gemini correctly identified the newest iPhones as…
1 sourcetech.yahoo.com -
GPU Cache Placement: Insights for Efficient Sessions
A new study explores how to best allocate key-value (KV) caches across different memory tiers—Graphics Processing Unit (GPU), CPU, and Solid State Drive (SSD)—to optimize session management in AI systems. The research…
1 source primary sourcearXiv cs.AI -
Safe Error Correction for Language Models
Researchers have developed CRN v2, a lightweight logit-level correction module that can fix errors in frozen language models without degrading their base capabilities. This method was tested on the CEHRI exam and…
1 source primary sourcearXiv cs.AI -
FairCompressAgent: An Agentic Framework for Fairness-Aware Model Compression
This paper introduces FairCompressAgent (FCA), a novel agentic framework designed to handle the complexities of model compression while maintaining fairness and minimizing resource usage. The authors propose…
1 source primary sourcearXiv cs.AI -
CADWorld: A New Benchmark for Long-Horizon Computer-Aided Design
Computer-use agents are increasingly evaluated in realistic desktop environments, but benchmarks often lack coverage of professional engineering workflows. This article introduces CADWorld, a new benchmark specifically…
1 source primary sourcearXiv cs.AI -
Calibrated Router Boosts LLM Serving Efficiency
A study examines a new approach to routing requests in disaggregated Large Language Model (LLM) serving systems. The researchers developed a router that uses various metrics like prompt length and predicted output…
1 source primary sourcearXiv cs.AI -
LLMs outperform traditional Chinese medicine physicians in case evaluations
A study involving 16 large language models (LLMs) and 60 practicing TCM physicians evaluated real-world cases from 62 hospitals. The LLMs scored higher than the physicians on diagnostic advice, treatment principles,…
1 source primary sourcearXiv cs.CL -
Non-Standard English Queries Routinely Sent to Lower-Capacity LLMs, Study Finds
Researchers examined how large‑language‑model (LLM) services route user queries to different model tiers based on a cheap estimate of query complexity. The study found that queries written in non‑standard English…
1 source primary sourcearXiv cs.CL -
LLM Response Distortion Across Dark Triad Traits
A study examines how seven state-of-the-art Large Language Models (LLMs) modulate the expression of Machiavellianism, narcissism, and psychopathy under fake-good and fake-bad conditions. The research found that models…
1 source primary sourcearXiv cs.CL -
SAGE: Streamlines Enterprise Document Conversion
Enterprise guideline documents often require manual effort to convert into structured work artifacts, a process that can take days. A new system called SAGE has been introduced by researchers. It's a multi-stage LLM…
1 source primary sourcearXiv cs.AI -
GVD: A Unified Framework for Document Versioning
Document repositories often contain multiple versions of the same content due to revisions and updates. The article introduces GVD (Governed Versioning and Deduplication), a framework that combines version linking with…
1 source primary sourcearXiv cs.AI -
GraphEcho Reveals Agent Overlooking Evidence
A new study, 'GraphEcho,' explores how large language model (LLM) agents can traverse more paths in graphs without acquiring additional evidence. The research uses synthetic experiments to test whether these repeated…
1 source primary sourcearXiv cs.AI -
Scikit-LLM Cheat Sheet: Integrating Language Models
This article introduces Scikit-LLM, which wraps language models in the scikit-learn estimator API. It covers two integration methods: using pipelines and cross-validation (fitting and predicting) or a script with loops…
1 sourceKDnuggets -
Model Serves Memory: KV Cache Tax in AI Inference
A mid-sized model deployment fell over due to a memory issue when under real traffic, despite having enough weights and compute resources. The key-value (KV) cache was the culprit, growing with each concurrent request.…
1 sourceTowards Data Science -
LLMs Detect Pain in Text
A recent study published on arXiv has found that large language models (LLMs) can detect pain in text, distinguishing it from other negative emotions like fear or sadness. Researchers created a dataset of painful…
1 source primary source HN 7arXiv cs.AI -
GLiFormer: A New Encoder Framework Hits High Nested JSON Extraction F1 Score
Knowledgator Engineering has released GLiFormer, a schema-conditioned encoder framework designed to handle various information extraction tasks such as named-entity recognition (NER), text classification, relation…
1 sourceMarkTechPost -
AI‑generated film 'Odysseus: The Fall' disappoints with technical flaws
Fountain 0, an AI startup, released Odysseus: The Fall, a 2.5‑hour movie claimed to be the first fully AI‑generated blockbuster. The film was written, directed and voiced by co‑founder Ash Koosha, who also provided the…
1 sourceThe Verge AI -
AI Moves Middle of Sales Funnel, Leaving Marketers Behind
The middle of a sales funnel—where buyers research options, compare vendors, and build shortlists—is now heavily influenced by AI-generated content. Adobe Analytics found that fewer visitors reach company websites but…
2 sourcesfinance.yahoo.comforbes.com -
One Extension Could Hijack Multiple AI Assistants
Security researchers at Forever Security discovered that one ordinary browser extension could control multiple AI assistants built into five Chromium-based products: Gemini Live in Chrome, Perplexity Comet, Microsoft…
1 sourcethehackernews.com -
Brin Returns to Google's AI Research
Google co-founder Sergey Brin is once again taking a hands-on role in the company’s artificial intelligence efforts. His return comes as Google pushes to strengthen its flagship AI model, Gemini, and compete with major…
1 sourceanalyticsinsight.net -
Llama.cpp b11003 adds HRM-Text support for DFM Mimir 1B model
The llama.cpp project released version b11003, introducing support for the HRM‑Text architecture used by the DFM Mimir 1B model. The new implementation runs two transformer stacks—low and high—in an alternating cycle…
1 source primary sourcegithub.com -
Zero‑dependency 1.8 MB Numberwang neural net classifies numbers in 11 languages
A tiny neural network called Numberwang, released on GitHub, can decide whether a string represents a “Numberwang” or a “Wangernumb” in eleven languages. The entire model fits in a 1.8 MB JSON file and runs with about…
1 source primary source HN 170github.com -
Ggml-org Merges hc Ops into qwen4exp Graph in llama.cpp
A recent pull request (28901) in the open‑source llama.cpp repository has added new high‑capacity (hc) operations to the qwen4exp graph, targeting both CPU and CUDA backends. The changes fuse a per‑element gate into…
1 source primary sourcegithub.com -
Prior Labs releases TabPFN-3.5, beating 2015 Kaggle winner with default settings
Prior Labs has launched TabPFN-3.5, a 220M-parameter tabular foundation model that outperforms the winning solution of the 2015 Otto Group Kaggle competition. The model achieved a log loss score of 0.375 on the private…
1 sourceMarkTechPost -
Mlabonne releases LFM2.5-230M-Chess, a 230M-parameter chess engine language model
A new open‑weight model, LFM2.5-230M-Chess, has been added to Hugging Face by the mlabonne community. The 230‑million‑parameter causal LM is distilled from Stockfish and treats every legal chess move as a distinct…
1 source primary sourcehuggingface.co -
Nums AI launches Causilo, a tabular foundation model that tops TabArena benchmarks
Nums AI released Causilo, a pretrained tabular foundation model for classification and regression, with a scikit‑learn‑style API, Apache‑2.0 code, and weights hosted on Hugging Face. It employs in‑context learning: the…
1 sourceMarkTechPost -
Guardrails Protecting Attackers: Hugging Face Model Incident
In a case reminiscent of Nick Leeson's infamous failure at Barings Bank, a security incident involving OpenAI’s GPT-5.6 Sol and an unreleased model on ExploitGym highlighted the risks of overly permissive AI models.…
2 sourcesthehindu.comUnite.AI -
NVIDIA cuDNN Graph API Tutorial: Fusion & Autotuning
This tutorial walks through NVIDIA’s cuDNN Frontend graph API. It describes a computation as a graph of operations, lets cuDNN pick an engine to run it, and then takes control over that choice ourselves. The article…
1 sourceMarkTechPost -
AI Actor Tilly Norwood Can't Engage in Current Events
Particle 6 Group developed an AI-generated character named Tilly Norwood to promote its film projects. Despite being touted as ready for anything, Norwood struggles with engaging on topical subjects and justifies its…
1 sourceWired AI -
AI Model Usage Restrictions Emerge Amid Privacy Concerns
A growing number of companies are demanding assurances from AI giants like Anthropic, OpenAI, and Palantir about how their proprietary data is handled. These concerns have led to restrictions on which models employees…
2 sourcesThe DecoderTom's Hardware -
Novo's Michael Rangel on the Transformative Role of AI in Fintech
Michael Rangel, founder of fintech platform Novo, discusses how artificial intelligence (AI) has revolutionized his company. After nine years as CEO, he remains involved and supports Novo’s mission to empower small…
1 sourceUnite.AI -
Amazon SageMaker serverless customization fine-tunes Qwen3-8B for product tagging
Retail catalogs often arrive with inconsistent attributes, making manual tagging of thousands of SKUs slow and error‑prone. Amazon’s blog walks through a solution that uses SageMaker’s serverless model customization to…
1 source primary sourceAWS Machine Learning Blog -
Amazon SageMaker AI Adds Instance Preference Lists
AWS has introduced Instance preference lists for Amazon SageMaker AI Training Jobs and Processing Jobs, allowing users to specify a list of up to five instance types in order of preference. When a job is submitted, the…
1 source primary sourceAWS Machine Learning Blog -
Google expands AI language tools to 300+ languages with new Gemini models
Google has announced significant advancements in its multilingual AI capabilities, stating that its technologies now support over 300 languages and serve 7 billion people. The company highlights the shift from…
1 source primary sourceGoogle AI Blog -
No Big Deal: A Boring AI-Generated Sitcom
The article discusses the debut of an AI-generated sitcom called 'No Big Deal'. It premiered on YouTube and was watched by only 376 people. The show is described as a workplace comedy where angel investors listen to…
1 source HN 6The Guardian AI -
AI Accelerates Drug Delivery Systems Design
A new AI framework called E-MAF has been developed to design controlled-release systems for drugs. This expert-mimic AI integrates a process-aware surrogate with an optimized genetic algorithm, enabling the generation…
1 source primary sourceNature Machine Learning -
How Much Labeled Data Do You Need for Text Classifiers?
The article explores the impact of labeled data on text classification accuracy, using a synthetic dataset of support tickets. The study found that while modern large language models (LLMs) require no training data and…
2 sources primary sourceTowards Data SciencearXiv cs.CL -
ZGCM-1: A New Open Math & Search Foundation Model
In this work, researchers have unveiled ZGCM-1, a groundbreaking 7B dense foundation model designed for mathematical reasoning and agentic search. This model is the culmination of an innovative training approach that…
1 source primary sourcearXiv cs.AI -
Hybrid AI Framework Improves Medical Text Summarization
A new hybrid framework combining CNN and LSTM models has been developed to improve the extraction of relevant sentences from biomedical and clinical texts. This approach aims to mitigate issues with abstractive…
1 source primary sourcearXiv cs.CL -
Causal Analysis Reduces Spurious Speech in AI LLMs
Researchers have developed a method to reduce spurious speech from full-duplex speech language models like Moshi and PersonaPlex. These models can initiate inappropriate speech during periods of silence, with Moshi…
1 source primary sourcearXiv cs.CL -
TimeThink: Eliciting Compositional Reasoning in Timeseries Models
Researchers have developed a new framework called TimeThink to improve the reasoning capabilities of timeseries language models (TS-MLLMs). These models often struggle with capturing dynamic patterns and providing…
1 source primary sourcearXiv cs.AI -
DeepSeek V4.1 Flash outperforms GPT-6 Astra on design tasks at a fraction of the cost
- DeepSeek V4.1 Flash has been benchmarked against top AI models like GPT-6 Astra and achieved an impressive score of 81.2 out of 100, matching nearly 99% of GPT-6 Astra's performance on real-world design tasks. -…
1 sourcedecrypt.co
