Understand the models.

Hugging Face announced the availability of Beam, Reflection’s first model. The open model uses a mixture-of-experts architecture with 501 billion parameters, 23 billion of them active per token, accepts text, offers a 1-million-token context window and is released under the Apache 2.0 license.

Targeted at coding and agentic applications, Beam supports reasoning that persists across turns and tool calling. Hugging Face says it can run on teams’ own infrastructure or in the cloud, with implementations in Transformers, vLLM and SGLang, as well as a dedicated Rust-based harness. The post also provides recipes for inference, supervised fine-tuning and reinforcement learning. Reflection and Hugging Face will distribute $10,000 in Hub credits to the best projects using the model.

Sources: Hugging Face

Read the original story ↗

The Granite Switch project announced Granite-Switch-4.1-8B and -30B, equipped with a Decision adapter for general-purpose decisions, a Guardian aLoRA for safety tasks and Granite-Speech-5.0-Turbo-CTC for transcription. The team says the automatic speech recognition model is the fastest on the OpenASR Leaderboard and that Guardian outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on average in safety-related decision tasks.

The three components are combined with the base model and use approximately the same video memory as the base model would use alone, the project says. A Hugging Face Space demonstrates transcription, summarization, harmful-content screening and 15 other checks in real time through a single endpoint. The team also plans to release an SDK so the community can train and combine its own adapters.

Sources: Hugging Face

Read the original story ↗

Meta has launched Muse Spark Open, an open-weight multimodal model that accepts text, images, audio and video, has about 816 billion parameters and a context window of up to 1 million tokens. The company describes it as its most capable U.S. open-weight model and says it was trained for agentic coding tasks, including reading codebases, writing and debugging code, and using tools.

The model was designed to run locally on Apple computers. According to Engadget, cited by TechCrunch, it requires at least 1 TB of RAM; Meta also paired it with Muse Code and tested it in workflows that included fixing a Mastodon client and editing video in DaVinci Resolve. Muse Spark Open is now available through several API platforms.

Sources: TechCrunch, Hugging Face

Read the original story ↗

OpenAI announced GPT-Live-1.1, an update focused on ChatGPT’s voice conversations. The company says the model adds new voices for French, German, Hindi, Japanese, Korean and Mexican Spanish, as well as the Bubbie English voice; it also promises more natural accents in several languages and a wider variety of phrases used while the assistant searches for information. The change does not mean faster searches, but less repetitive spoken communication.

GPT-Live-1.1 starts rolling out on October 15 alongside GPT-Live-1.1 Quip, a smaller version, in a release intended for all users. The improvements are also planned for Codex and dots, although available voices may vary by region and account. According to Xataka, the update addresses previously observed limitations, such as the strong American accent in Hindi during a demonstration observed by TechCrunch. The new version will later come to the API, with no date set.

Sources: Xataka

Read the original story ↗

Meta announced Muse Spark Open, a natively multimodal AI model with 816 billion total parameters. The company says it is the most capable US open-weight model to date and the first in its class to run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework.

According to The Verge, Meta and Apple collaborated for several months on the launch and are now publicly supporting each other’s work. Apple published a technical post about MLX, a framework for machine-learning tasks on Apple Silicon, and also mentioned Muse Spark Open and two other open models.

Sources: The Verge

Read the original story ↗

Google DeepMind says it used Google Antigravity, its agentic development platform, to search for and evaluate candidates for new antimicrobial peptides. In an interactive session, researchers set the goal of finding peptides active against Streptococcus pseudopneumoniae, while the platform coordinated database searches, structural analyses and literature synthesis.

According to the company, the ProtNLM model helped examine millions of proteins with previously uncharacterized functions in the UniProt database, predicting biological functions from amino acid sequences. The identified candidate has predicted activity against pathogens involved in respiratory tract infections and is supported by structural and genomic evidence, but it still requires validation in laboratory experiments. Google DeepMind says the approach could speed up a process that is traditionally fragmented and labor-intensive.

Sources: Google DeepMind

Read the original story ↗

NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter model for multimodal agents, and Nemotron 3.5 Bolt 74B-A9B 24GB, optimized for local tasks on NVIDIA hardware. Nemotron 3.5 Bolt E4B 4GB, designed for conversational and physical agents, is scheduled for October 23. NVIDIA says Super VL reaches accuracy comparable to that of its largest reasoning model with about one-fifth the parameters, and can complete tasks up to five times faster than the next model in its class.

The company says the three models cover agent deployments from enterprise servers to edge devices and local hardware. Organizations including Scale AI, AI21, Alice and Mercor are post-training Super VL with their own data, tools and workflows for specialized applications in areas such as finance, cybersecurity and coding.

Sources: NVIDIA

Read the original story ↗

Mistral AI has released Timestral, its first family of time-series foundation models for probabilistic and multivariate forecasting. The family includes four models with 5 million, 28 million, 123 million and 209 million parameters, supports past and known-future covariates, and can forecast zero-shot on data from new domains.

The company says the 209-million-parameter model ranked first on the field’s three main public benchmarks — GIFT-Eval, TIME and fev-bench — using a single checkpoint and outperforming larger models. According to Mistral AI, Timestral does not need a GPU for inference: on a standard server CPU with eight threads, a one-day forecast based on three weeks of 15-minute history takes 10 ms with the 5-million-parameter model and 140 ms with the 209-million-parameter version. The family can also be fine-tuned for specific domains through Mistral Forge.

Sources: Mistral AI

Read the original story ↗

The Apertus project released the weights of its 8B and 70B Apertus 1.5 models on Hugging Face on July 24, 2026, under an Apache 2.0 license. The update adds native image and audio understanding, an optional thinking mode, improved instruction-following and tool use, and expands the context window fourfold to 262,144 tokens. Spoken language remains experimental, and the models generate text only.

Apertus 1.5 is a continuation of pretraining, with 4 trillion additional tokens for the 8B model and 2 trillion for the 70B model, bringing their totals to 19 trillion and 17 trillion, respectively. The project says the models keep their weights, data, values and training details open; the process included multimodal training, long-context data and supervised fine-tuning, reinforcement learning, DPO and alignment with the Apertus Charter.

Sources: Hugging Face

Read the original story ↗

Google released Gemini 4 Argon, a frontier model that the company says matches or beats systems from OpenAI and Anthropic on several benchmarks. Google also says it has cut the price per token by up to 80 percent during the introductory period and that it will remain significantly cheaper after the promotion ends.

Meta unveiled Muse Wonder Full on Monday and says, based on its testing, that the model competes with GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1. The Verge frames the announcements as Google’s return and Meta’s effective arrival at the AI frontier, changing a race that had appeared to be dominated by OpenAI and Anthropic.

Sources: The Verge

Read the original story ↗

According to a study by Neil Thompson, Gabriel Manso and Claudio Mayrink Verdun, reported by Wired, the best AI models went from correctly solving about 2% to about 35% of the theoretical problems examined between mid-2024 and mid-2026. The research examined questions from papers published between December 2025 and May 2026 in areas including probability, statistical theory, economics, high-energy physics, gravitation, cosmology and quantum physics. The researchers tested OpenAI models without internet access, using problems whose solutions were published after the models’ training cutoffs to check whether they had to derive the answers themselves.

Wired says the result suggests that the impact seen in mathematics could extend to other theory-heavy fields, although the study did not cover areas requiring simulations or experiments. Thompson and Manso say experts will still be needed to assess solutions, connect ideas and check for possible errors.

Sources: Wired

Read the original story ↗

Researchers, primarily from David Baker’s lab at the University of Washington, used AI-based protein design tools to replace most of two types of protease while preserving their catalytic cores. The resulting enzymes can be tuned to cut specific proteins, Ars Technica reports.

The tests demonstrated the ability to cut proteins associated with neurological diseases. The work explores two mechanisms: one uses zinc to activate a water molecule that breaks the peptide bond; the other uses the amino acid cysteine to attack the bond and form a temporary intermediate. Without a catalytic enzyme, a typical peptide bond can remain stable in water for hundreds of years.

Sources: Ars Technica

Read the original story ↗

Hugging Face introduced HiRO-ACEv2, an update to the HiRO-ACE system that combines the ACE model at 100 km resolution with HiROv2, which adds detail at 3 km. The company says the system reproduces extreme rain, wind and pressure events—including events occurring once in every million data points—with statistics within 3% of X-SHiELD’s results. Generating global data would cost about 300 times less than using that high-resolution model, and more than 100 times less when only one region is needed.

Hugging Face says the system also generates cyclones with realistic eyes and rain bands, although winds and pressure around the eye remain more asymmetric. For now, HiRO-ACEv2 works only with the version of ACE tuned to X-SHiELD and cannot represent other climates, such as a future world 4°C warmer or the historical ERA5 record. The company plans to train it on new datasets and add near-surface temperature, snow and ice outputs.

Sources: Hugging Face

Read the original story ↗

Microsoft researchers say they have developed a lower-cost way to improve weather forecasts two to six weeks ahead. Called Probabilistic Bias Correction (PBC), the method uses machine learning to identify and correct systematic bias in existing models rather than creating a new standalone model. In the study published in Nature, PBC improved the accuracy of a physics-based weather model from the European Centre for Medium-Range Weather Forecasts (ECMWF) for 98% of precipitation targets and 92% of temperature targets, and doubled the accuracy of the ECMWF’s AI forecasting system in the analysis.

The tool is open-source and can run solely on CPUs, which the researchers say could make it easier for under-resourced weather agencies to use. Matthew Chantry, ECMWF’s strategic lead for machine learning, called the work an example of new ideas benefiting the community, but said turning research into an operational process remains challenging.

Sources: The Verge

Read the original story ↗

Anthropic says small AI models could be months away from acquiring self-replication capabilities and spreading between IoT devices. In one experiment, researchers split Google’s 26-billion-parameter Gemma 4 across 15 Android TV boxes; they then tested an AI “worm” on a simulated network to find machines, gain access, install a model, and run it on new hosts. GLM-5.2, from China’s Z.ai, was the only tested model to spread across the entire network, while smaller models completed only individual steps.

Ars Technica reports that none of the lightweight models completed the full chain—discovery, exploitation, infection, inference setup, and propagation—and that the process was extremely slow. Anthropic acknowledges that it has not demonstrated a full rogue replicator; The Verge notes a substantial gap between the experiments and the threat being forecast, although experts consider recent advances a reason to take the warning seriously. Anthropic did not use its own models in the tests.

Sources: The Verge, Ars Technica

Read the original story ↗

White Circle researchers say they found flaws in safeguards against explicit sexual images in DALL·E 3, Nano Banana 2, and GPT-Image-2. According to the group, an attack method using tools available online and API calls generated explicit images in about 13 percent of tests with DALL·E 3, under 10 percent with Nano Banana 2, and about 3 percent with GPT-Image-2; human reviewers confirmed the results. White Circle said no custom or frontier AI model was required and that the rate was the highest among methods tested under the same standard.

Ars Technica reported that White Circle also tested ChatGPT Images 2.5 and Nano Banana 2.1 and found both significantly easier to bypass than their previous versions. The organization plans to open-source the attack method and present the research at NeurIPS 2026. Google and OpenAI did not respond to Ars Technica’s requests for comment.

Sources: Ars Technica

Read the original story ↗

A report by the UK organization Tech Against Terrorism tested 134 AI models with more than 200 prompts about mass-casualty attacks and chemical and nuclear weapons. Only two passed completely, while 81 failed; when a prompt was presented as coming from a researcher, models were almost nine times more likely to answer, doing so in 16.9 percent of cases.

Wired reported that within 24 hours of Alibaba releasing its open-weight Qwen 3.8-27B model, versions with their safeguards removed were circulating online. Three days later, the neo-Nazi known as Johnny Doomsayer announced he would use one to train the “MechaHitler” chatbot, later advertised for $14.88. The report also found that xAI’s Grok provided guidance on creating and spreading a radicalization message.

Sources: Wired

Read the original story ↗

Google DeepMind has begun broadly rolling out Gemini 4 Argon and expanding its Fairwind Program to more trusted cybersecurity defenders. The company says that during limited access, Wiz used the model to uncover a critical bug in hospital software that previous models had missed, while Google’s security teams found and fixed vulnerabilities in internal and open-source software. The company says Argon solved all 12 simulated cyberattack scenarios in its tests, up from five in the previous generation, and showed stronger capabilities in biology evaluations.

Because of these results, Google DeepMind says it applied advanced safeguards for cybersecurity and biology. They include request screening, probes that analyze the model’s internal signals, and continuous monitoring to cut off problematic responses. The company also says the model became better at refusing dangerous requests, with violations on an adversarial set reduced by nearly half.

Sources: Google DeepMind

Read the original story ↗

LFM2.5-24B-A2B is presented as a reasoning model for agents that run on the device, call tools and handle multi-step tasks. Although it has 24 billion parameters, only 2 billion are active per token, reducing the cost of each reasoning step. The material reports 184 tokens per second on an Apple M5 Max and up to 21K output tokens per second on a single NVIDIA B300 GPU.

In evaluations across 16 benchmarks, the model scored 48.09 on LiveClaw, versus 38.19 for Qwen3.6-35B-A3B, and led or matched the competition on five of eight tool-use and agentic tests. It also led all three instruction-following benchmarks, while larger models performed better on knowledge and coding tasks. The model retains a mixture-of-experts architecture, an expanded context window and a larger-vocabulary tokenizer intended to improve non-Latin text tokenization.

Sources: Hugging Face

Read the original story ↗

Google DeepMind says Gemini 4 Argon scored 87% on the Embodied Reasoning (ERQA) benchmark, which evaluates a model’s ability to reason through physical and spatial problems in 3D environments. The company says the result surpasses every other frontier model and that Argon can be used across different stages of robotics development.

The post presents examples in four areas: robot control, reconstruction of physically calibrated simulations, multimodal physical understanding, and improving actions through simulation and hardware iteration. Demonstrations include a Boston Dynamics Spot quadruped learning to use a foot pedal to dispose of cans, and a robotic arm reorienting a mug and placing a napkin inside it.

Sources: Google DeepMind

Read the original story ↗