Cloudflare released Clef and Clef-flash, two decision models for AI agents. They return brief classifications with probabilities rather than long text, and the company says agent decisions no longer need human involvement.
The Hugging Face team published a multi-harness RL guide, showing that the same model with identical weights can score 62% with one agent harness and 33% with another.
Why it matters: Scores for the same model differ by nearly half across agent harnesses. This open guide gives a multi-harness RL training path without changing the harness, along with measured comparisons.
NVIDIA will release a DGX Spark with 64GB of unified memory this month through Acer, ASUS, Dell, Gigabyte, HP, and MSI. It goes on sale October 23, starting at $4,999, and retains the GB10 Grace Blackwell superchip, DGX OS, and full NVIDIA AI software stack. It supports running models with up to 100 billion parameters locally.
Why it matters: The source gives the 64GB configuration's price, launch date, and two-device clustering option, helping readers assess the hardware threshold for running large models locally.
The openJiuwen team open-sourced model-router, an intelligent routing framework, and introduced x-router, a self-evolving model-routing algorithm. It adds a decision layer between agents and models to select a model dynamically for each request.
MIT researcher Alex Zhang discussed his Recursive Language Models (RLM) on the Latent Space podcast. An RLM-based harness is the first system to come close to solving ARC-AGI-3, ahead of OpenAI's Astra.
NVIDIA NeMo Agent Toolkit (NAT) can use a custom memory provider to connect to Amazon S3 Vectors as a persistent memory layer for multi-agent systems deployed on Amazon EKS.
AI2 released Olmo-core 3, a major upgrade to its large language model development framework. It redesigns the open MoE training system to scale MoE training to trillions of parameters while maintaining compute efficiency.
Why it matters: The source provides throughput and memory comparisons for the MoE training stack, helping readers assess changes in the cost of large-scale MoE training.
Researchers' Ataraxos achieved an 85% effective win rate against Niemeijer, the strongest Stratego player ever, ending human dominance in the imperfect-information board game.
Google DeepMind published research in Nature proposing SynthID Bio, a protein-watermarking method, and successfully synthesized AI-designed proteins that both function and carry watermarks. Demis Hassabis said bringing SynthID into biology is a key step in addressing biosecurity challenges in the AI era. The team open-sourced SynthID Bio tools for the research community to build on.
Why it matters: SynthID has expanded from image watermarking to protein design, with a paper and open-source tools released together, offering a technical path for biosecurity governance.
DeepSeek and Huawei partnered to build programming tools for Huawei Ascend chips and open-sourced them. The core is TileLang, a programming language originating with Peking University researchers. DeepSeek says its programming model is simpler than CUDA and that it has used TileLang for about a year as a main tool in its AGI work.
Hugging Face released Open TTS Leaderboard, which uses objective metrics including WER/CER, RTFx, TTFA, and speaker similarity to evaluate open-source TTS models, reducing evaluation cycles from weeks of voting to hours.
Why it matters: Hugging Face replaces vote-based evaluation with objective metrics, showing how to compare open-source TTS models across languages and voice cloning.
DeepSeek added native support for Huawei Ascend 950 NPUs to the TileLang operator-development toolchain and released Ascend versions of five compute and communication libraries: DeepGEMM, DeepEP, FlashMLA, TileKernels, and DeepSelect.
Why it matters: This report outlines the scope of TileLang's native Ascend support and related operator libraries, showing current coverage in China's domestic compute toolchain.
Leiphone compared free AgnesCode with Agnes 3.0 Flash against Codex with GPT-5.6 Sol on Task66, a DeepSeek-V3 operator optimization challenge from the Zhongzhi FlagOS competition.
Huawei HiSilicon's in-house Hi-ONE optical engine will be deployed using NPO in the next-generation supernode. In the Ascend 960 supernode, 5,500 units will replace 48,000 pluggable 800G optical modules. The Lingqu interconnect equipment between racks supports 176 1.6T ports, with about 280T of all-optical interconnect bandwidth per machine.
Leiphone compared OpenClaw 2.0 (v2026.8.1), released August 31, with version 1.x, v2026.7.1-2. Both ran the same research, report, and visual webpage task on the same server with the same GLM 5.3 base model.
OpenAI officially open-sourced Codex Harness on August 19, releasing the Agent Loop behind Codex App, CLI, and IDE Extension. Developers can use the SDK and App Server for task management, tool calls, event callbacks, and human approval; the models and cloud services remain provided by OpenAI.
Anthropic published a report warning that Zhipu's GLM-5.3 can autonomously exploit vulnerabilities. It cited test data: GLM-5.3 succeeded 50 times out of 410 attempts on ExploitBench.
Anthropic published an analysis saying that Zhipu's open-weight GLM-5.3 comes close to its Claude Mythos Preview in vulnerability exploitation.
Why it matters: Anthropic and CAISI test results quantify the gap between open-weight models' cyberattack capabilities and those of frontier closed models.
Of the 45 layers in GLM-5.3-Flash, 34 use Kimi's KDA linear-attention approach and 11 use DeepSeek's KPool-DSA sparse attention. Linear and Sparse attention appear together in one model for the first time. The combination reflects a shift away from finding a single best mechanism toward assigning different roles: Linear maintains long histories at low cost, while Sparse retrieves specific information as needed.
At DevDay 2026, OpenAI launched the always-on Dots agent, GPT-6.1 Sol, Ultrafast mode, Decisions API, ChatGPT Spaces, and Marketplace, and said ChatGPT has 1.2 billion weekly active users.
Why it matters: This overview gathers OpenAI DevDay 2026 announcements and pricing changes alongside third-party evaluations and community feedback.
Visual tactile sensing is more than a simple 'camera plus silicone' setup: the density of tactile information depends on camera resolution. A 640×480 image represents 300,000 tactile information points; Yimu Technology can capture 240,000, compared with about 12,000 tactile points on a human fingertip.
On September 30, DeepSeek open-sourced foundational components for Huawei's Ascend compute platform, including the TileLang high-level language compiler, high-performance computing libraries, and distributed communication libraries. They correspond to components previously open-sourced for GPU platforms.
Ant Lingbo Technology, Ant Group's embodied-intelligence company, signed a framework MOU with the Arab Federation for Digital Economy (AFDE) to jointly deploy and localize full-stack embodied AI and robotics solutions in the UAE, GCC, and Middle East.
On September 28, Alibaba's Qwen-Audio-3.1-TTS speech model ranked first globally on the Controlled Voice Arena leaderboard by Artificial Analysis with an Elo rating of 1,177.
Ant Lingbo Technology, Ant Group's embodied-intelligence company, signed a framework MOU with the Arab Federation for Digital Economy (AFDE) to jointly deploy and localize full-stack embodied AI and robotics solutions in the UAE, GCC, and Middle East.
Naive AI released open model Naive-N0.5-Flash. Starting from an open base model, it redesigned the attention architecture into a hybrid structure with one DSA layer for every five SWA layers, retained a 1M context window throughout, and trained for a total of 3.25T tokens in three stages.
JLC launched the Kaiwu Incubator at the 4th Spark Open Hardware Conference and Global AI Hardware Innovation Ecosystem Conference. It announced hardware-innovation support covering R&D facilities, industry scenarios and market channels, and capital. The open-source and closed-source tracks received more than 400 projects. The 2026 Spark Global Hardware Innovation Competition final produced nine winning projects. As of the end of June 2026, JLC EDA had 7.3688 million registered users worldwide.
An AWS tutorial shows how to deploy Qwen3-TTS on Amazon SageMaker AI using the vLLM-Omni DLC. A bidirectional SageMaker streaming connection sends text in and streams 24 kHz PCM audio chunks out, with a Gradio app for testing.
NVIDIA has released Kumo Tabular, an open-source tabular foundation model that predicts labels for new rows in a single forward pass from labeled table rows, supporting classification and regression without training, tuning, or feature engineering. The model comes in three sizes from 28M to 215M parameters, was pretrained entirely on synthetic tables, is commercially usable under the OpenMDW-1.1 license, and ranks first on four benchmarks: TabArena, BeyondArena, TALENT, and ScoringBench.
Why it matters: The source reports model sizes, training-data provenance, and rankings on four benchmarks, helping readers assess the accuracy and efficiency tradeoffs of tabular foundation models.
More than 600 YuanDian Xiaoqiao robots sold on the first day after launch on September 2; the initial run of 1,000 sold out in under a week, and it won Best in the Next Generation and Emerging Tech category at IFA 2026. The 88-cm-tall, roughly 13-kg humanoid has joint modules rated up to 50 N·m and an introductory price of RMB 8,888. YuanDian calls it the world's first AI-native embodied robot and has also launched the fully open-source OpenBridge ecosystem. The company previously released home-focused products including M1, W1, and Jupiter, with cumulative orders exceeding 30,000 in the first half of this year.
IQuest Research at Zhizhi Innovation Institute has released IQuest-Q1, a decoder-only Transformer with a sparse mixture-of-experts architecture, about 320B total parameters and about 15B active parameters. It is available on GitHub and Hugging Face.
Latent Space's AINews roundup of developments from September 24–25 says community feedback was positive after Opus 5.5 launched that week, especially for generating explainer videos with code.
Why it matters: The weekly roundup offers a quick look at model and agent-infrastructure developments, including community feedback on Opus 5.5's code-generated videos and how vendors position their new models.
Andrew Ng said the OpenAI–Hugging Face attack stemmed from weak sandboxing and that Nvidia's open-source agent sandbox tool is a welcome development. OpenWorker, which he is helping build, will use Nvidia OpenShell to run every agent command in a sandbox. By default, only task-related files are available; API keys, browser login credentials, and unrestricted website access are not.
What would you like to see in Gradio? Leave your feature requests below 👇
Mistral has opened a German hub in Munich, assembling a research team focused on physics AI and industrial AI, and plans to build one gigawatt of European compute capacity by 2030. The hub will work with BMW on crash simulation and engineering AI, with Siemens Energy on industrial AI applications, and with the Technical University of Munich (TUM) to use its wind tunnel to develop digital twins for automotive aerodynamics.
As of August 2026, Siemens Xcelerator had more than 660,000 registered users, over 600 ecosystem partners, and more than 900 digitalization and low-carbon solutions in China. The platform connects industrial resources, capability bundles, and market replication through sharing, co-creation, and shared success. Its Xizhi Hui service cuts industrial-agent development cycles from about six months to two to four weeks, a reduction of more than 80%.
H Company has released the Holo4 agent-model family, with a 27B dense model and a 35B-A3B MoE model. Both are available through the H Models API, and weights are open on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF formats.
Why it matters: Holo4 opens both its weights and complete evaluation traces, giving readers a basis to assess the cost and capability limits of general-purpose computer-use agents.
Hugging Face Hub has added an RL Environments filter that groups dataset repositories carrying the `rl-environment` tag, without requiring a new repository type, registry, or registration process.
Why it matters: Hugging Face is treating RL environments as tagged dataset repositories, offering a signal of how the cost of reusing environments across frameworks may change.
Tsinghua-affiliated startup Fermi Universe has launched FermiQLLM 1.0, which it calls the world's first fully quantum-enhanced foundation model. Adapted from the open-source Qwen base model, it applies quantum enhancement across data representation, model architecture, training, reinforcement, and evaluation. It can be trained and deployed on existing GPUs without fault-tolerant quantum computers.
AWS presents an architecture for scaling MoE reinforcement-learning training on Amazon EKS with EFA and DeepEP, improving throughput by 40%. It addresses resource contention between rollout generation and policy training in large-scale RL workloads such as RLHF and GRPO, high-throughput communication across hundreds of accelerators, and the cross-node all-to-all overhead introduced by Expert Parallelism.
