AI model news
Stories 301 to 360 of 419, from 17 September 2026 to 25 September 2026, newest first. The latest are on the first page →
-
Alibaba releases Qwen-Image-2.1 with transparent background support
Alibaba's Qwen team released Qwen-Image-2.1 on September 20, 2026. This image generation model combines creation and editing in a single system, allowing users to generate images with transparent backgrounds directly.…
1 sourcenote.com -
Liquid AI releases LFM2.5-VL-DSpark for faster vision-language model inference
Liquid AI announced LFM2.5-VL-DSpark, a draft model for its LFM2.5-VL-3B vision-language model. The model uses speculative decoding to speed up inference by up to 3.13x on-device and 2.66x on H100 GPUs, with end-to-end…
2 sources primary sourceMarkTechPostHugging Face -
Microsoft adds GPT-6 Sol, Luna and Astra to Foundry model lineup
Microsoft’s chief technology officer Mark Russinovich warned that today’s most critical single points of failure are often AI models or inference endpoints rather than traditional infrastructure. He said the usual…
2 sourcesnote.commsdynamicsworld.com -
Google DeepMind chief Koray Kavukcuoglu says Gemini 4 is in post-training and aims to release it much earlier than end of 2026
Google DeepMind CEO Koray Kavukcuoglu confirmed in media appearances that Gemini 4 has entered the post-training stage and that the company aims to release an early version 'as soon as possible', targeting a launch…
3 sourcestimesnownews.comThe DecoderThe Verge AI -
Anthropic engineer says Claude's writing declined as newer models focus on code
Jackson Kernion, an Anthropic engineer who works on Claude fine‑tuning, said the model’s writing quality has stalled or even worsened despite improvements in math, code, and reasoning. He explained that after Opus 4.6…
2 sourcesnote.comThe Decoder -
AI expert warns machine-speed cyberthreats may outpace human defenses
Yagub Rahimov, CEO of Polygraf AI, warns that AI-driven cyberattacks are evolving into coordinated ‘swarms’ of autonomous agents capable of operating at machine speed. Unlike traditional bots or agents, these swarms…
1 sourcefoxbusiness.com -
OpenAI may unveil GPT-6 Cyber cybersecurity model in weeks
OpenAI is preparing to unveil GPT-6 Cyber, its fourth cybersecurity‑focused large language model of 2026, according to a Fortune report. The model is slated to arrive in the coming weeks and is designed to help…
5 sourcestimesofindia.indiatimes.comthenews.com.pknote.com -
Fastino Releases GLiNER2.5-Decide: A 340M Open-Weight Decision Model That Runs on CPU
The model takes text and a schema of typed questions and returns answers with probability distributions, confidence scores, and constraint-feasibility metadata. It is built on a DeBERTa-v3-large encoder, fine-tuned…
1 sourceMarkTechPost -
Pistis introduces 27B- and 9B-parameter multimodal models via new training framework
Researchers released the Pistis model family on arXiv, featuring two multimodal large language models built on Qwen3.6 and Qwen3.5. The 27B-parameter model uses Qwen3.6, while the 9B-parameter model is based on…
1 source primary sourcearXiv cs.AI -
OpenAI releases ChatGPT Images 2.5 with sketch, point, and template tools
OpenAI has rolled out ChatGPT Images 2.5, a new image‑generation update that lets users sketch ideas, point to parts of an image, edit photos, use templates, and share prompts. The update focuses on better image…
4 sourcesnote.comenglish.mathrubhumi.com -
FreedomIntelligence releases HuatuoGPT-3 multimodal model on Hugging Face
The model builds on Qwen3.8-27B with a reinforcement-learning method called One-stage Policy Optimization (OnePO). It requires a thinking mode during inference, generating reasoning before final answers, and supports…
1 source primary sourcehuggingface.co -
PrismML launches 1-bit Bonsai LLM for Qualcomm smart glasses
PrismML, a lab founded by Caltech researchers and advised by UC Berkeley’s Ion Stoica, unveiled a version of its tiny language models that can run on Qualcomm‑powered smart glasses. At Qualcomm’s Snapdragon Summit on…
1 sourceTechCrunch AI -
BottleCap AI cuts Qwen3.8-27B’s reasoning tokens by 37.2% with ThinkingCap-Qwen3.8-27B
BottleCap AI released ThinkingCap-Qwen3.8-27B, a fine-tuned version of Qwen’s Qwen3.8-27B designed to reduce reasoning tokens by 37.2% across 12 benchmarks. The model maintains near-identical accuracy (a 0.86pp drop to…
1 sourceMarkTechPost -
What has changed with GPT-6 Astra? Basic specifications of the new model that supports 'thinking, researching, and creating' in ChatGPT
Unlike previous models, Astra organizes information, makes judgments, and chains processes like reading a document, researching gaps, and compiling a report into a single workflow. It includes a reasoning.effort…
2 sourcesnote.com -
GPT-6 Astra and Claude Opus 5.5 compared on pixel art game task
A content creator tested two frontier AI models, GPT-6 Astra and Claude Opus 5.5, by giving them the exact same prompt to generate a browser-based game. The task required creating a Japanese-style pixel art game…
1 sourcenote.com -
Model adds Ling 3.0 VL support
The pull request 29151 merged Ling 3.0 VL support into the model on September 19, 2026. Contributors aetherbird, CISCreviewed, and ngxson reviewed and approved the changes. The update folds Ling 3.0 VL into the…
1 source primary sourcegithub.com -
Contrastive-LM releases CLM-8B, an open model that scores actions up to 9× faster than Jev
Contrastive-LM has released CLM-8B, the first open Contrastive Language Model, which scores candidate actions instead of generating text. The model uses a frozen Qwen3-8B backbone with trainable projection heads and is…
2 sourcesMarkTechPostcontrastive-lm.notion.site -
Xiaomi releases MiMo-V2.6-Pro-RL, a 1.02T/42B open omnimodal model trained via scaled RL
Xiaomi has released the MiMo-V2.6-Pro-RL model, a flagship open-weight omnimodal AI model with 1.02 trillion total parameters and 42 billion activated parameters, supporting text, image, video, and audio modalities…
6 sources primary sourcenote.comThe DecoderLatent Spacehuggingface.co -
Claude edges ChatGPT and Gemini in 1000-character story test
A user conducted a creative writing experiment asking ChatGPT, Claude, and Gemini to write a complete short story within 1000 characters. The prompt required the AI to choose its own genre and theme, avoid safe…
1 sourcenote.com -
Anthropic has a 71% chance of leading AI model market by end of 2026
Market participants assign Anthropic a 71 % probability of holding the leading AI model by the end of 2026, based on data from prediction‑market platform Kalshi. The odds follow the company’s recent rollout of Claude…
1 sourcecryptobriefing.com -
Google Antigravity SDK adds support for local Gemma 4 26B models
Google has announced that the Antigravity SDK now supports local AI workflows, enabling developers to run agentic capabilities entirely offline. The initial release features support for the Gemma 4 26B A4B model using…
1 sourcedevelopers.googleblog.com -
Nvidia releases Nemotron 3 Diarization, a 100M-parameter model tracking up to 8 speakers
Nvidia announced the open‑weight speaker diarization model Nemotron 3 Diarization on Hugging Face. The 100 million‑parameter model can identify who spoke when in audio streams, handling up to eight overlapping speakers…
1 sourceMarkTechPost -
Anthropic gives Claude users one free usage reset token
Anthropic announced the release of Claude Opus 5.5 on September 22, 2026, alongside an increase to the 5-hour usage limits for paid plans such as Pro, Max, and Team. As part of this update, the company distributed a…
1 sourcenote.com -
Alibaba's Qwen-Image-2.1 claims to beat closed models in image generation
Alibaba's Qwen AI team released Qwen-Image-2.1, an open‑weight image generation and editing model with 7 billion parameters. The team says it outperforms most closed models on its own benchmark, though independent…
3 sourcesTom's HardwareMarkTechPostThe Decoder -
Alibaba launches Qwen Audio 3.1 with five speech models and price cuts
Alibaba’s Qwen AI team released Qwen Audio 3.1, a suite of five models covering speech‑to‑text (ASR), an enhanced ASR‑Next, text‑to‑speech (TTS), a diffusion‑based TTS‑Next, and a real‑time interaction model. The ASR…
1 sourceThe Decoder -
Kyutai releases Voice of Reason speech-to-speech model for spoken math
Kyutai announced two open‑weight speech‑to‑speech models, called Voice of Reason, built on the GLM‑4‑Voice‑9B foundation. The models solve math problems directly from spoken input, removing the need for a transcription…
1 sourceMarkTechPost -
Researchers unveil Ovis-Embedding, a universal omni-modal embedding family
A team of researchers announced Ovis-Embedding, an omni‑modal embedding family that encodes text, images, video and audio in a single representation space. The system builds on a pretrained Qwen‑omni backbone, which is…
1 source primary sourcearXiv cs.AI -
PAANI delivers on-device visual evidence fusion for river robots
PAANI is an on‑device perception‑to‑guidance system designed for mobile river monitoring robots. It fuses a YOLO11n object detector and a MobileNetV3 Small semantic segmenter on an Arduino UNO Q, producing explainable…
1 source primary sourcearXiv cs.AI -
OpenAI's GPT-6 Astra deciphers 108-year-old WW1 German radio cipher
OpenAI released GPT-6 Astra on September 3, 2026. The flagship model, already noted for playing Minecraft autonomously and building nested simulations, was used by AI developer Prinz to break a 108‑year‑old German…
1 sourcedexerto.com -
AI Safety Concerns Erupt Amid Recursive Self-Improvement Fears
This week, fears over AI safety and the potential for AI to cause human extinction gained new traction. Evan Hubinger, an alignment lead at Anthropic, expressed concern that AI could kill all humans within the next…
4 sourcesThe DecoderInterconnectsthehindubusinessline.comCNBC Technology -
Aws preview TBC rat‑Brain video model for select customers
Starting Tuesday, Amazon Web Services will let a limited set of enterprise customers test The Biological Computing Company’s (TBC) rat‑brain AI model, which the startup says can speed video generation up to five times…
2 sourcesThe DecoderWired AI -
ChatGPT-6 Astra decodes 108-year-old WWI German cipher
A developer known as Prinz claims to have used ChatGPT-6 Astra to break a 108‑year‑old World War I German radio message that was listed among 50 unsolved ciphers on Scienceblogs.de. The cipher, identified as using the…
2 sources HN 50Tom's Hardwarecryptocellar.org -
JS Denain of Epoch AI says Chinese models lag US labs by about six to eight months
In a podcast with Nathan Lambert, Epoch AI senior researcher Jean‑Stanislas “JS” Denain discussed the state of AI acceleration and the US‑China capability gap. He noted OpenAI’s blog showing a 2× per‑month increase in…
1 sourceInterconnects -
Tencent launches Hy Image 3.5 preview, says it matches ByteDance's Seedream 5.0 Pro
Tencent Holdings Ltd. announced the preview of its new image‑generation model, Hy Image 3.5. The company says internal testing showed the model’s output was on par with ByteDance’s Seedream 5.0 Pro, a leading…
1 sourcebloomberg.com -
Hugging Face adds GGUF support to Transformers for local inference
Hugging Face announced that the Transformers library can now load GGUF‑format checkpoints directly, letting users run quantized models on a laptop with the familiar frompretrained API. The feature reuses llama.cpp’s…
1 source primary sourceHugging Face -
Rumors of new AI models may surface next week
The author, tk, shares rumors that “incredible new models might be released one after another by various companies next week.” He lists rumored models such as Claude Opus 5.5 (or 5.2), Kimi 3.1, Gemini 4 (Gemini 4…
1 sourcenote.com -
Austrian Academy of Science to release Apollo, first advanced Greek language model
The Austrian Academy of Science announced that on Wednesday it will release Apollo, the world’s first advanced large language model for Ancient Greek. Developed with French AI lab Mistral and technology services firm…
1 sourceWired AI -
Opinion: users complain about preachy ChatGPT, OpenAI tweaks GPT‑5.3 Instant
Users overseas are voicing frustration that ChatGPT often sounds like a lecturer rather than a helper, a sentiment that has been spreading on Reddit and other forums. The complaint centers on unsolicited warnings and…
1 sourcenote.com -
xAI launches Grok 4.7 for coding and knowledge work at $2 input and $6 output tokens
xAI has launched Grok 4.7, its newest model for coding and knowledge work, priced at $2 per million input tokens and $6 per million output tokens. The model uses a larger base model than Grok 4.6, features longer…
5 sources primary sourceMarkTechPostAWS Machine Learning BlogThe DecoderUnite.AI +1 more -
Claude cannot forecast crypto or stock prices, benchmark shows humans outperform
Anthropic’s Claude model does not have live market data and cannot predict cryptocurrency or stock prices. A peer‑reviewed benchmark presented at ICLR 2025 found that expert human forecasters beat the best large…
2 sourcesnote.comhackernoon.com -
OpenAI classifies GPT-6 Astra as Critical for cybersecurity
OpenAI released GPT-6 Astra on September 3 as a limited preview, then added it to paid ChatGPT tiers a day later. The company’s president said the model might mark the start of the AGI era, but the most notable detail…
1 sourceTowards Data Science -
StepFun launches Step 5 Preview with 1M-token context and $1 input pricing
StepFun announced Step 5 Preview, its flagship mixture‑of‑experts model aimed at agentic software engineering, knowledge work, and finance tasks. The model totals about 600 billion parameters but activates roughly 27…
1 sourceMarkTechPost -
mini-AGI is a continual learning byte-level model that trains on a single 8 GB GPU
mini-AGI is a byte‑level language model that builds its own architecture and trains from scratch on a single 8 GB VRAM GPU. It pages weight files from disk so the parameter count is limited by free disk space rather…
1 source primary sourcegithub.com -
Tencent unveils Gander model that talks while handling background tasks
Tencent's Hunyuan Speech team introduced Gander, a model that can keep a real‑time conversation going while a separate "brain" component works on complex tasks. The architecture splits responsibilities: a fast…
1 sourceThe Decoder -
Runway says it wants real-time AI video generation as a live stream
Runway presented research on turning AI video generation into a live, controllable stream. The company says users would describe a scene and see video frames appear instantly, instead of waiting for a completed clip.…
1 sourceThe Decoder -
Alibaba releases Qwen3.8-Omni-Flash, a 1M-token omni-modal model with agentic audio‑video
Alibaba’s Qwen team announced the launch of Qwen3.8-Omni-Flash, its first omni‑modal model that handles text, images, audio and video and returns text. The model runs as a hosted API on QwenCloud, Alibaba Cloud Model…
3 sourcesMarkTechPostThe Decoder -
This looks promising: stepfun-ai/Step-5-Preview-BF16 · Hugging Face
this looks promising: stepfun-ai/Step-5-Preview-BF16 · Hugging Face Hugging Face Models Datasets Spaces Buckets new Docs Enterprise Pricing Website Tasks HuggingChat Collections Languages Organizations Community Blog…
1 source primary sourcehuggingface.co -
Anthropic says Claude now leads 26% of its AI research and development
Anthropic announced that its Claude model is now responsible for roughly 26% of the company’s model research and development work, and it participates in about 90% of R&D tasks under human supervision. The firm said…
12 sourcesmoneycontrol.comnote.comyahoo.comtechxplore.com +4 more -
Anthropic's Invisible Text Watermarking in Claude AI
Anthropic has introduced a watermarking system for their Claude AI text generation tool to identify AI-generated content more reliably. This system subtly adjusts word choices during text generation, creating a…
5 sourcesnote.comUnite.AIArs Technica AIcryptobriefing.com +1 more -
NASA and IBM release open-Source Lunar Foundation AI Model
NASA and IBM Research announced the open‑source NASA‑IBM Lunar Foundation Model, a multimodal AI system pretrained from scratch on SomBench, a lunar dataset containing nearly two million co‑registered data bundles…
1 source HN 42newsroom.usra.edu -
OpenAI founder launches Jev model promising lower cost and faster processing
A founder of OpenAI has announced a new AI model called Jev, which the developer says offers the same performance as existing large language models while costing significantly less and delivering faster response times.…
2 sourcesnote.com -
Linkup Research releases sparseup 149M-parameter sparse embedding model
Linkup Research announced SPARSEUP, an open‑source sparse embedding model built on a 149 million‑parameter ModernBERT backbone. The model is released under the Apache 2.0 license on Hugging Face and can be loaded via…
1 sourceMarkTechPost -
Six Jev-inspired models released in two days, from Laya (421M params) to Kev-0.5B
Jev’s launch video attracted 36M views in two days, far behind OpenAI’s Navier Stokes result (74M) and Anthropic’s Fable 5 (57M). Within the same period, six independent clones appeared, each positioning itself as a…
1 sourceLatent Space -
MarkTechPost explains GGUF, GPTQ, AWQ, EXL2, EXL3 model formats
The article separates two layers of model handling: containers that store tensors on disk and quantization methods that compress weights. Containers mentioned are safetensors, GGUF and the older PyTorch pickle files…
1 sourceMarkTechPost -
xAI releases Grok Voice Transcribe 2.0 speech-to-text model
On September 18, 2026, xAI introduced Grok Voice Transcribe 2.0, its newest speech‑to‑text model. The service keeps batch pricing at $0.10 per hour of audio and $0.20 per hour for streaming, and xAI says the model is…
2 sourcesMarkTechPostUnite.AI -
Alibaba open-sources Damo Radar model to detect 146 abdominal conditions
Alibaba Group Holding's research arm, Damo Academy, has open-sourced Damo Radar, a vision-language AI model designed to identify nearly 150 abdominal conditions from contrast-enhanced CT scans. The model covers 18…
1 sourcescmp.com -
Jina AI releases jina-ocr-v1 3.4B MoE document parser for low‑budget GPUs
Jina AI, part of Elastic, announced the open‑weight release of jina-ocr-v1, a visual document parser that converts PDFs, scans, tables, charts or invoices into clean Markdown in a single pass. The model contains 3.4 B…
1 sourceMarkTechPost -
World model firms keep product plans secret, executives say
At the All In conference, a panel on world models highlighted the secrecy surrounding the field’s leading companies. Yann LeCun’s AMI Labs and Fei‑Fei Li’s World Labs have attracted buzz and funding but have not…
1 sourceTechCrunch AI -
PrismML hopes its tiny LLM will change how we all use AI
PrismML hopes its tiny LLM will change how we all use AI If AI lab PrismML isn't on your radar yet, it should be. If AI lab PrismML isn’t on your radar yet, it should be — not because it’s raised gobs of money (it…
2 sourcesMarkTechPostTechCrunch AI -
Moonshot AI's Kimi K3 opens on Amazon Bedrock with 2.8 trillion parameters
Moonshot AI announced that its Kimi K3 model is now available on Amazon Bedrock. The model is described as the first open‑weight model to reach 2.8 trillion parameters and adds native vision support with a…
2 sources primary sourceUnite.AIAWS Machine Learning Blog
