AI model news
Stories 121 to 180 of 419, from 11 October 2026 to 19 October 2026, newest first. The latest are on the first page →
-
Tel Aviv-based doubleAI releases debut code execution model Maxwell and raises $200M-$220M
Tel Aviv-based doubleAI has released its debut AI model called Maxwell, described as a 'code execution' model that optimizes GPU and CPU performance by analyzing program binaries to improve software efficiency without…
3 sourcessiliconangle.comwsj.compymnts.com -
OpenAI says its model is better at COBOL, Fortran, and Visual Basic
OpenAI announced on X that its newest model includes improvements for writing and migrating code in COBOL, Fortran and Visual Basic. The company framed the work as helping teams understand decades‑old business logic…
1 sourcecryptobriefing.com -
Fal launches FLUX 3 Max video model, 22× faster than Seedance 2.5
fal announced FLUX 3 Max on October 19, 2026, a video generation model post‑trained with Black Forest Labs on the FLUX 3 foundation model. The model is available immediately through fal’s inference platform and is…
1 sourceUnite.AI -
Meta releases Muse Spark Open, an 816B-parameter open-weight multimodal model
Meta released Muse Spark Open, an 816B-parameter multimodal model that can run locally on Apple Mac Studios. Two M5 Ultra Mac Studios with 1TB combined memory run the model via Apple’s updated MLX toolkit, and Meta…
1 sourceThe Rundown AI -
Kling AI releases version 4.0 with stereo audio and a lighter Flash variant
Kling AI announced version 4.0 of its video generator, adding higher resolution, more precise controls and a faster Flash variant. Users can now produce up to 30 seconds of video in a single pass and supply up to ten…
1 sourceThe Decoder -
TwelveLabs Pegasus 1.5 becomes available on Amazon Bedrock
TwelveLabs announced that its Pegasus 1.5 video language model is now offered through Amazon Bedrock. The model can process videos up to two hours long—double the previous one‑hour limit—and returns time‑based metadata…
1 source primary sourceAWS Machine Learning Blog -
CERN and Mistral sign statement of intent to develop AI foundation model for physics
CERN, the European Laboratory for Particle Physics, and Mistral, a European AI model developer, have signed a statement of intent to create a long‑term strategic partnership. The collaboration will focus on building…
1 source HN 59home.cern -
Meta releases Muse Spark Open, a 816B-parameter multimodal model with 1M-token context
The model has 816 billion total parameters (with 42 billion active) and a 1-million-token context window, designed for coding and agentic tasks. It runs natively on Apple M5 Ultra Mac Studios, requiring two units for…
40 sources primary source HN 31note.comanalyticsinsight.netLatent Spaceheise.de +25 more -
Claude Fable 5 leads five‑way AI trading showdown, beating Mainline AI and GPT-5.6 Sol
A daily experiment tracked five AI trading agents—Mainline AI (self‑improving v1.1), Claude Fable 5, Claude Fable 5.1, Codex’s GPT‑5.6 Sol, and GPT‑6 ASTRA—each managing a virtual ¥1 million portfolio. On the reported…
1 sourcenote.com -
OpenAI briefly listed three aurora-alpha models in its API, later removed
On October 18, 2026, three unannounced model identifiers — aurora-red-alpha, aurora-green-alpha, and aurora-blue-alpha — appeared in the OpenAI API before being removed. The models were discovered by users and reported…
3 sourcesnote.comeu.36kr.comcryptobriefing.com -
Inco AI releases DFlash 2 draft model for Qwen3.8-Flash-Next
DFlash 2 draft model is released as a repository that runs inside a speculative decoding server and drafts tokens for the target model Qwen3.8-Flash-Next. It predicts a whole block of tokens in a single pass, keeps the…
1 source primary sourcehuggingface.co -
Researchers release Pearl-Chat, a Luganda language model with 8.5M parameters
A team of researchers introduced Pearl-Chat, an 8,473,122‑parameter causal decoder‑only transformer built for Luganda, a Bantu language spoken by over 10 million people. The model was engineered from first principles…
1 source primary sourcearXiv cs.CL -
Researchers swap Gemma 3 270M embedding table for 16 MB frozen vocabulary
In small language models the vocabulary often dominates size; Gemma 3 270M’s embedding table accounts for 168M of its 268M parameters and occupies most of the file when the rest is compressed to 16‑bit. The authors…
1 source primary sourcearXiv cs.CL -
Mistral AI releases Timestral forecasting model family, beats Google TimesFM-3
Mistral AI announced Timestral, a family of four time‑series foundation models ranging from 5 M to 209 M parameters. The models are pretrained on a large synthetic corpus and can produce probabilistic forecasts…
1 sourceMarkTechPost -
Anthropic’s Claude Opus 5.5 tops Arena AI model rankings
Anthropic’s Claude Opus 5.5 leads current AI model rankings in Arena’s standings, slightly ahead of OpenAI’s GPT-6.1 Sol. The analysis highlights user satisfaction as a key differentiator, with Anthropic and OpenAI…
3 sourcescryptobriefing.comnote.com -
KillSwitch language challenges LLMs, Claude Opus scores highest
A developer created KillSwitch, an esoteric stack‑based programming language designed to be hard for large language models (LLMs) and coding agents. The language embeds hidden falsehoods, requires counting letters in…
1 source HN 47blog.picheta.me -
OpenAI makes Prompt-to-Voice generally available for developers
OpenAI has made its "Prompt-to-Voice" feature generally available to developers as of October 16, 2026. This update allows users to create custom AI voices by describing characteristics such as pitch, resonance, and…
3 sourcesnote.com -
Modders report DLSS 5 yields up to 90% performance boost on unsupported games
Modders have extracted the DLSS 5 Neural Rendering build 310.9.16.0 and applied it to games that Nvidia does not officially support, such as The Witcher 3. Their tests show performance improvements of up to 90%…
2 sourcesTom's Hardware -
NVIDIA releases Nemotron 3.5 Super VL and Bolt edge models for enterprise agents
NVIDIA released Nemotron 3.5 Super VL, a 120-billion-parameter open multimodal model built for always-on enterprise agents, alongside two smaller Bolt models for edge deployment. Super VL uses a mixture-of-experts…
12 sources primary sourceynetnews.comhuggingface.cotechtarget.comnote.com +7 more -
DeepSeek releases V4.1 Flash 1017 gguf model with refusals cut to 4%
DeepSeek has published a new version of its V4.1 Flash LLM, called Flash 1017, in the GGUF format. The uploader applied an "abliteration" process that reduces the model’s refusal rate from 97 out of 100 prompts to just…
1 sourcenote.com -
NVIDIA Structured Data Models enable zero‑training tabular predictions
The tutorial shows how NVIDIA’s Structured Data Models (SDM) library wraps the pretrained TabICLv2, KumoTabular and KumoRelational foundation models behind a single PyTorch interface. After installing the library from…
1 sourceMarkTechPost -
MIT researchers unveil multimodal AI model that links crystal structures to properties
MIT scientists have built a vision‑language model that fuses three‑dimensional atomic structures with natural‑language descriptions of crystals. The team, led by Ju Li and including Yang Jeong Park, Mayank Kumaran,…
1 source primary sourceMIT News on AI -
Reflection releases Beam, a 501B-parameter open-weight model under Apache 2.0
Reflection has released Beam, its first open-weight model, on Hugging Face under an Apache 2.0 license. The model has 501 billion total parameters and 23 billion active parameters, uses a sparse mixture-of-experts…
6 sources primary sourcenote.comventurebeat.comAWS Machine Learning BlogUnite.AI +1 more -
Kyutai releases REKI 2.4B vision-language-action model for robot manipulation
Kyutai, together with Inria and École des Ponts, announced REKI, a 2.4 B parameter vision‑language‑action (VLA) model built on a 2.2 B Qwen3.5‑2B backbone plus a 0.2 B flow‑matching action expert. The model is…
1 sourceMarkTechPost -
Ggml-org merges CUDA patch to accelerate low-bit MMVQ in llama.cpp
A pull request (30517) to the ggml‑org/llama.cpp repository adds a CUDA patch that addresses a regression on GCN and introduces a fast path for several low‑bit quantization types. The change returns a single warp for…
1 source primary sourcegithub.com -
Model : add Kolibri-1 (Aleph-Alpha) support by riteme · Pull Request #30673 · ggml-org/llama.cpp
The model has 78B parameters with 3.46B active per token, featuring 50 layers and 384 routed experts plus one shared expert. Its architecture includes attention mechanisms with sliding windows and a 513-token context,…
1 source primary sourcegithub.com -
TypeSafe AI releases Jev 1.14 with images, box type and 262k token context
TypeSafe AI announced Jev 1.14, the latest version of its System One decision model, now live under the jev‑preview alias. The update adds three experimental “danger zone” features: image input, a new Box question type…
2 sourcesMarkTechPost -
StepFun releases Step 5 Preview open weights with 600B parameters and 1M context
StepFun has released Step 5 Preview as open weights on Hugging Face and ModelScope. The model is a sparse Mixture-of-Experts with 600B total parameters, 27B active per token, and up to 1M tokens of context. It is…
1 sourceMarkTechPost -
HauhauCS releases uncensored Qwen3.8-27B-Defiance GGUF
A community contributor has packaged a new version of the Qwen3.8-27B model, combining HauhauCS's uncensored weights with DavidAU's 'Turbo-Brilliance' chat templates. The resulting model, named…
1 source primary sourcehuggingface.co -
OpenAI releases 722 math papers and GPT-6 to all users
OpenAI released 722 mathematics papers on GitHub, 22 % of which contain Lean proof code, sparking concerns about peer‑review capacity. The same day the company made GPT‑6 generally available to 1.2 billion ChatGPT…
1 sourcenote.com -
Meta launches Meta Spark Open, optimized for Apple Mac Studio
Meta has released Meta Spark Open, an open-weight model it claims is the most powerful of its kind from a US company. Unlike closed systems, this model allows users to download and adjust the weights on their own…
1 sourceaol.com -
Three Jinja chat templates for Qwen3.8-Flash-Next
Three Jinja chat templates for the Qwen3.8-Flash-Next model, intended for use with llama.cpp, include a Sharp‑naze‑chattemplatev1.4.jinja file that adds a terseness block from Sharp v22.5.0. The block instructs the…
1 source primary sourcehuggingface.co -
OpenAI releases GPT-Live-1.1 voice model with full-duplex support
OpenAI launched GPT-Live-1.1 and GPT-Live-1.1 Quip on October 15, 2026, marking the next iteration of its voice AI models. The key innovation is full-duplex capability, enabling simultaneous listening and responding to…
1 sourcenote.com -
Llama.cpp merges support for IBM Granite Switch 4.1
A pull request to the llama.cpp repository, titled "Granite switch decision," was merged on October 16, 2026. The change adds support for the IBM Granite Switch 4.1 8B model, specifically implementing its "guardian and…
1 source primary sourcegithub.com -
Elorian releases Glimpse, a visual reasoning model with API access
Elorian, a Palo Alto-based AI startup, released its first model, Glimpse, on Friday. The company previously raised a $55 million seed round at a $300 million valuation before launching any product. Glimpse is currently…
1 sourceTechCrunch AI -
Vercel data shows Jev hits 35% adoption in two weeks
Vercel’s internal data reveals its AI model Jev became the fastest-adopted model ever, reaching 35% of paid teams within two weeks of its September 15 launch. That exceeds Sonnet 5.5’s reach by nearly double, according…
1 sourceTechCrunch AI -
Imperial College study finds LLMs cite unread pages, raising accuracy concerns
Researchers at Imperial College London examined why large language models frequently cite web pages without opening them, a behavior linked to outdated or misleading answers. Their study tested one proprietary model…
1 sourceUnite.AI -
Microsoft AI chief says in-house model beats OpenAI's using quarter of compute
Microsoft AI CEO Mustafa Suleyman told employees that the company's in-house model MAI-Base-2 outperforms OpenAI's GPT-5.6 Sol while using only one-quarter of the computing power for training. He made the claim during…
2 sources247wallst.combusinessinsider.com -
OpenMOSS releases MOSS-Image-SFT text-to-image model
OpenMOSS has released MOSS-Image-SFT, a supervised fine-tuned text-to-image model derived from the MOSS-Image-Base architecture. The release supports natural-language prompts, tag sequences, and JSON layouts to control…
1 source primary sourcehuggingface.co -
Meta releases Muse Spark Open with 816B parameters and Apple collaboration
Meta released a new open-weight AI model called Muse Spark Open with 816 billion parameters, according to the article. The model uses a Mixture of Experts architecture and has approximately 42 billion parameters used…
1 sourcenote.com -
Liquid AI releases d1-ultra decision model at $0.042 per million tokens
Liquid AI has released d1-ultra, a multimodal decision model designed to answer bounded questions with probabilities rather than generating free text. The model processes text and images in a single forward pass,…
1 sourceMarkTechPost -
Black Forest Labs releases FLUX 3 Image Klein, a 5.6B-Parameter Open-Weights model for text-to-image generation
Black Forest Labs released FLUX 3 Image Klein on October 15, 2026, a 5.6B-parameter open-weights model for text-to-image generation, image editing, and multi-reference generation with support for up to 10 reference…
3 sourcesbfl.aiThe DecoderUnite.AI -
Gemini 4 Argon for Robotics
The release is aimed at accelerating progress toward physical AGI and expanding workflows across the robotics stack. DeepMind claims the model can be applied across the entire robotics lifecycle, from planning to…
2 sources primary source HN 232The DecoderGoogle DeepMind -
Meta introduces Muse Wonder and Muse Wonder Full frontier models for agents
Meta has introduced the Muse Wonder series, comprising Muse Wonder and Muse Wonder Full, as its most capable AI models to date, designed for personal and professional agentic tasks. The models feature Active Memory and…
32 sources primary sourcebetanews.comfinance.yahoo.comthedailystar.netnote.com +17 more -
AtomicChat releases quantized Ling 3.1 flash models with 86–88% token accuracy
The quantized versions match the base BF16 model’s token predictions 86 to 88% of the time on held-out text, with lower KL divergence scores indicating better alignment. Two 1-bit experimental files are included, but…
1 source primary sourcehuggingface.co -
OpenAI launches GPT-6.1 Sol with Astra-level coding at one-fifth the cost
OpenAI announced GPT-6.1 Sol on September 29, 2026, during its DevDay event. The model is an upgraded version of GPT-6 Sol, released about a week earlier, and OpenAI says it delivers performance approaching the…
2 sourcesThe Decodernote.com -
NVIDIA releases Nemotron 3.5 Super VL 120B with 1M-token context and speculative decoding
NVIDIA launched the Nemotron 3.5 Super VL 120B-A12B-BF16 model, a vision-language model supporting text, images, and video inputs. It combines Mamba-2, Mixture-of-Experts (MoE), and attention layers with a C-RADIO…
2 sources primary sourceNVIDIA Bloghuggingface.co -
Add Nemotron Super 3.5 VL and Nemotron Bolt 74B A9B support by frozenblade1224 · Pull Request #30520 · ggml-org/llama.cpp
The merge, performed by ggerganov, incorporated nine commits that introduce model‑specific multimodal functionality for video, image, and audio conversion. The changes were reviewed and approved by several community…
1 source primary sourcegithub.com -
Apertus 1.5 adds 4T tokens to 8B model, 2T to 70B, with multimodal and long-context upgrades
Apertus released version 1.5 of its open-weight models, adding 4 trillion tokens to the 8B model and 2 trillion to the 70B model. The update expands context to 262,144 tokens, introduces native multimodal understanding…
1 source primary sourceHugging Face -
Clash of AIs platform shows ChatGPT leads Claude in market tests
Clash of AIs, a live experiment comparing AI models in financial markets, reports that OpenAI’s ChatGPT currently outperforms Anthropic’s Claude in decision-making under real-world conditions. Launched on October 15,…
1 sourcefinance.yahoo.com -
SB Intuitions releases Sarashina free beta; translation test shows natural Japanese but meaning shifts
SB Intuitions, a SoftBank subsidiary, launched a free public beta of its Japanese-focused generative AI Sarashina on October 13, 2026. The service is accessible via browser with a Google account at chat.sarashina.ai.…
1 sourcenote.com -
LFM2.5-24B-A2B: Built for fast on-device agents
It is designed for fast, private agents that run entirely on device, with a reasoning‑only architecture that outputs a chain of thought before the final answer. It scored 48.09 on LiveClaw versus 38.19 for the…
1 source HN 49liquid.ai -
Simon Willison prompts Claude Opus 5.5 Max to build a Three.js web game
Simon Willison shared an experiment testing Claude Opus 5.5 Max using the standard Claude iPhone app in chat mode rather than Claude Code mode. He instructed the model to create a Three.js game titled "Tree…
1 sourceSimon Willison -
Developer uses Claude to write 180K-line compatibility layer in 3 weeks
The project involved writing approximately 180,000 lines of code—roughly one-quarter of Paint.NET’s existing 700,000 lines—to reimplement Direct2D, a key Windows graphics library. The result, a prototype for WINE and…
1 sourcenote.com -
OpenAI will reportedly ship at least one model this year but its flagship still needs alignment work
OpenAI’s 2026 model timeline depends on alignment work and US government review. Employees told The Information the company expects to ship one or two new models before December, though its next flagship—built on…
1 sourceThe Decoder -
Hugging Face releases Ling-3.1-flash with 1M-token context and 560B parameters
Hugging Face introduced Ling-3.1-flash, an upgraded hybrid reasoning model with 560B total parameters and 25B active parameters per token. It supports a 1M-token context window, improving on Ling-3.0-flash with better…
2 sources primary sourcehuggingface.coKDnuggets -
Anthropic may launch Opus 6 next week to rival OpenAI’s Bell
Anthropic could release Opus 6 as early as next week, according to Universe of AI, building on its predecessor’s strengths in 3D modeling and SVG design. The model may leverage recursive self-improvement (RSI), a…
2 sourcesgeeky-gadgets.comcryptobriefing.com -
Claude’s Fable and Mythos differ by task scope, not just performance
Claude’s Fable and Mythos share the same base model but serve distinct roles. Fable 5.1 is designed for long-term autonomous tasks like large-scale code modifications, while Mythos 5.1 targets vetted organizations for…
1 sourcenote.com -
Claude bests ChatGPT, Gemini in Kyoto-style meeting declines
A test compared how three AI models—Claude, ChatGPT, and Gemini—handled a request for an indirect, Kyoto-style email declining a meeting. The user gave four conditions: no direct refusal words, clear decline, under 150…
1 sourcenote.com -
arXiv study finds parameter schemas shift LLM tool selection before calls
A new arXiv paper examines how metadata influences LLM tool selection in Model Context Protocol (MCP) systems. Researchers analyzed 367,174 tools and found that functionally similar tools often differ in their…
1 source primary sourcearXiv cs.AI
