CATEGORY
Open source
Open models, weights, tooling, datasets and community infrastructure.
Google and Meta open-source native PyTorch support for TPUsTorchTPU lets existing PyTorch workloads run on Google TPUs with minimal code changes and supports training, serving and deployment tools including vLLM and SGLang. The release is intended to make TPUs a more practical alternative to GPU-centric AI software stacks.→DeepSeek open-sources its production inference engineDeepInfer supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, with optimisations for scheduling, speculative decoding and large-model execution. The release does not include DeepSeek’s full cache-management, prefill/decode-disaggregation or load-balancing stack.→SGLang 0.5.22 delivers faster model loading and a critical state-correctness fixSGLang 0.5.22 introduces an InstantTensor loader that maintainers say makes models load up to ten times faster, adds new model and diffusion support, and improves DeepSeek-V4.1 throughput on H200 GPUs. It also fixes a bug that could resume hybrid-linear models from incorrect Mamba states when speculative decoding and structured output were combined.→Meta releases 816-billion-parameter open-weight Muse Spark OpenThe natively multimodal mixture-of-experts model has 816 billion total parameters, 42 billion active parameters and a one-million-token context window. Meta positions it for coding, long-horizon agentic work and multimodal reasoning, with weights available through major model hubs and inference providers.→Valeo and KE:SAI launch an open autonomous-driving research collaborationThe collaboration will use reinforcement learning in simulation and trained world models to expose systems to rare and safety-critical driving situations. The partners said they plan to release the code, trained models and datasets as open source.→Mistral opens public preview of trillion-parameter Large 4 modelMistral Large 4 is a natively multimodal mixture-of-experts model with about 1.05 trillion total parameters and 52 billion active parameters. The preview supports a one-million-token context window, while model weights are planned for later release.→OVERT becomes an open stewardship model for verifiable AI safeguardsOVERT specifies records that can show whether declared governance, security and oversight controls were applied when an AI system acted, including the tool calls made by agents. The specification is voluntary and does not certify safety or establish legal compliance, but its source text and implementation guidance are publicly available.→Alibaba releases open-weight Qwen Image 2.1Qwen Image 2.1 is available through Hugging Face and ModelScope for visual generation and editing tasks. The model supports transparent-image generation and can use up to ten reference images, under a research and evaluation licence.→Open Secure AI Alliance moves under Linux Foundation governanceThe Linux Foundation announced that the Open Secure AI Alliance had joined its ecosystem under neutral governance. The alliance plans open tools, shared standards and a findings exchange spanning models, agents, identity, policy, containment and infrastructure.→Hugging Face launches Open Alignment InitiativeClement Delangue said the initiative would be led by Hugging Face co-founder Thomas Wolf. The organisation said alignment work should not remain behind the closed doors of a small number of frontier labs.→Motional open-sources nuReasoning dataset for autonomous-driving edge casesThe nuReasoning dataset was developed with UCLA’s Mobility Lab and focuses on common-sense reasoning for difficult driving situations. Motional said an earlier miniset had already surpassed 50,000 downloads and that the release is intended to support driving foundation-model research.→Nvidia agrees to buy Hugging Face for $12.93BNvidia said it will acquire Hugging Face for $12.93B, including a $1B employee retention plan. Hugging Face serves more than 18 million developers, researchers and creators, and Nvidia said it will remain an open platform supporting multicloud and multi-accelerator development.→Apodex 1.1 pairs multi-agent reasoning with built-in verificationApodex announced version 1.1 as a system for professional, scientific and financial work that combines multi-agent collaboration with built-in verification. The release also includes Apodex 1.1 Mini as open weights and FrontierAgent as open-source software, although the company’s benchmark claims remain vendor-reported.→Tencent releases Hy4 preview, a 770-billion-parameter open-source modelHy4 preview has 770 billion total parameters and approximately 49 billion active parameters per request, according to Tencent. The model is open-sourced and available through Tencent products and APIs, although Tencent cautions that the early release may over-verify answers and take longer on complex tasks.→Visa expands its open-source security harness from discovery to validated remediationThe latest VVAH release adds closed-loop remediation, validation, flexible model selection and real-time workflow visibility. Visa says some vulnerability-resolution cycles have fallen from weeks to hours, while its documentation still requires human review of generated findings and fixes.→Perturb launches a decentralised network for continuous AI-model red-teamingThe network supports black-box, white-box and transfer attacks, returning robustness scores, vulnerability heatmaps and hardening datasets to model owners. Perturb says findings will be privately disclosed and that the system is intended to provide continuous external testing rather than periodic red-team engagements.→NVIDIA reportedly nears $12.9 billion Hugging Face acquisitionThe Information reported that NVIDIA agreed to acquire Hugging Face for $12.9 billion, while another report said the talks were not signed and could still collapse. Neither company confirmed the transaction to TechCrunch.→Perceptron releases open-weight Isaac 0.5 industrial vision modelPerceptron launched Isaac 0.5 as an open-weight model for vision-guided robots working in settings such as warehouses and factories, combining perception, reasoning and action in one system. The company says it trained the model with about one million hours of general video, ego and UMI recordings, and petabyte-scale multimodal and robotic-trajectory data.→IBM releases Granite 4.2 open reasoning models for enterprise agentsGranite 4.2 is available in 3B, 8B and 30B sizes, with IBM describing native thinking and agentic reinforcement learning for multi-step enterprise tasks. The release also includes lightweight Granite Speech 5.0 Turbo CTC models intended for fast transcription on edge devices.→Marin begins a publicly tracked 535B-parameter open training runThe announced run uses 80% pretraining and 20% midtraining across an 18.75T-token mixture over roughly three months, with estimated compute of 2.7e24 FLOPs. Marin says its data mixture, sampled documents and live training loss will be available through GitHub and Weights & Biases.→SGLang and AntLing cut trillion-parameter restart time to 32 secondsAntLingAGI and Alibaba reported 0.63-second weight loading for the Ling-2.6-1T FP8 model, up to 780 times faster than disk loading. The end-to-end engine restart fell to about 32 seconds by avoiding repeated disk reads.→TIER IV joins OIN 2.0 to reduce patent risk for autonomous-driving softwareTIER IV said OIN 2.0 participation will reduce patent risk for automakers, suppliers and semiconductor companies building on Autoware and related open-source software. The move extends open-source governance beyond code licensing into patent protection for vehicle compute and autonomy ecosystems.→HiFi-UMI makes robot-free demonstrations competitive with teleoperation in tested tasksThe report evaluates HiFi-UMI across three policy backbones and four bimanual tabletop tasks, finding success differences of -2.5, +3.1 and -0.6 percentage points against in-domain real-robot teleoperation baselines. It also reports 85% success on a precision-insertion task and lower offline action-prediction error with pretraining, while limiting the conclusions to the tested models, tasks and conditions.→TIER IV and Renesas pair Autoware with automotive AI compute for software-defined vehiclesTIER IV plans to optimise Autoware and its end-to-end models on R-Car Gen 5, targeting solutions from Level 2+/2++ driver assistance to Level 4 autonomy. The companies plan a demonstration on R-Car X5H at Automotive World 2026; the announcement is a collaboration roadmap rather than a production deployment.→Open-weight AI faces a new US government scrutiny debateThe current US AI framework excludes released open models from its definition of covered frontier models. Axios reported discussion of pre-release review, national-security evaluation and procurement standards for powerful open-weight systems.→Z.ai turns its cyber-capable models towards vulnerability discoveryZ.ai says its GLM family has been used to identify more than 2,400 vulnerabilities, including more than 1,000 rated critical or high. The figures are vendor-reported, but the initiative connects frontier-model cyber capability with defensive security work.→Z.ai opens vulnerability scanning to open-source maintainersThe programme is linked to Z.ai's GLM model family and its new OpenVuln initiative. Z.ai said its models have identified more than 2,400 vulnerabilities, including flaws rated critical or high.→Open-weight AI faces a new US government scrutiny debateThe current US AI framework excludes released open models from its definition of covered frontier models. Axios reported that officials and industry participants are nevertheless discussing pre-release review, national-security evaluation and procurement standards for powerful open-weight systems.→