Startup Arcee AI has released and open-sourced its 400-billion-parameter large language model Trinity, claiming it was trained over six months using approximately 2,048 Nvidia Blackwell B300 GPUs at a total cost of about $20 million. The model is released under the Apache license, emphasizing 'true open source' to differentiate from restrictive licensing models. Currently text-only, versions with vision and speech capabilities are in development. The company has also released smaller variants, Trinity Mini and Nano, for free download, and plans to launch hosted APIs with pricing details within approximately six weeks.

Google has started rolling out its AI agent feature Auto Browse in Chrome, based on Gemini 3 and Project Mariner, capable of automatically performing web tasks such as finding information and filling out forms. It launches via a sidebar panel, runs in new tabs, and supports parallel multitasking. The feature is currently available only to AI Pro and AI Ultra subscribers: the former get 20 uses per day, the latter 200. Operations are executed in the cloud, with page content temporarily logged to users' Google accounts, though the system will not automatically complete high-risk actions like purchases.

Tencent has announced the open-sourcing of its Hunyuan Image 3.0 image-to-image model, based on a Mixture-of-Experts (MoE) architecture with a total of 80 billion parameters and around 13 billion active parameters. The model introduces a 'think-before-edit' workflow that performs image understanding and reasoning prior to editing, improving instruction following and local edit precision. Official benchmarks show the model ranks among the top-tier globally on the LMArena image editing leaderboard (blind evaluation dimensions include instruction adherence and fidelity). Tencent claims Hunyuan has spawned approximately 3,000 image- and video-related models, promoting an open ecosystem for generative content production.

Pillar Security has disclosed the 'Bizarre Bazaar' operation, where attackers profit by exploiting exposed or weakly authenticated LLM service endpoints. Over 40 days, more than 35,000 attack sessions were recorded. Attackers commonly scan for unauthenticated Ollama port 11434, OpenAI-compatible API port 8000, and publicly accessible MCP servers, then verify access and promote paid services named silver[.]inc on Telegram and Discord, accepting payments via cryptocurrency or PayPal. Risks include compute theft, API reselling, exposure of prompts and conversation data, and potential lateral movement within internal networks via MCP.

AI data labeling company Handshake has acquired data quality firm Cleanlab, primarily a talent acquisition, with Cleanlab’s nine core employees—including three MIT PhD co-founders—joining Handshake’s research team. Founded in 2021, Cleanlab specializes in automatically identifying mislabeled data to improve training set quality and had raised approximately $30 million in funding. Originally a recruitment platform, Handshake has recently entered the AI data labeling space, reportedly serving top labs including OpenAI. This acquisition highlights the strategic importance of training data quality and validation capabilities in the large model supply chain.

The Allen Institute for AI (AI2) has released and open-sourced SERA (Soft-verified Efficient Repository Agents), a series of coding agent models targeting independent developers and small-to-medium enterprises. The 32B version solves 54.2% of tasks on SWE-bench Verified, while the 8B variant achieves 29.4%. AI2 states it collaborated with Nvidia to optimize interfaces for improved production readiness, and has open-sourced the models, training methods, and integration approaches with Anthropic’s Claude Code, supporting 'one-line code launch.' A cost comparison shows that reproducing top open-source model performance costs about $12,000, while using SERA costs approximately $400.

Google has launched TranslateGemma, a set of open-source weight translation models based on the Gemma 3 architecture, offering 4B, 12B, and 27B parameter sizes covering up to 55 languages, with emphasis on support for low-resource languages. The models target different deployment scenarios: 4B for mobile/edge devices, 12B for laptops and local research, and 27B for high-fidelity cloud applications. Sources indicate the models are available for download on Kaggle and Hugging Face, and can be deployed via Vertex AI. Despite no dedicated multimodal training, they can still handle text-in-image translation tasks.

Ant Group's Lingbo team has open-sourced LingBot-VLA, a foundation model for embodied intelligence trained on approximately 20,000 hours of real-world operational data, emphasizing validation of Scaling Laws in embodied AI through real-world data scale. The model adopts a 'Brain + Cerebellum' MoE Transformer architecture, combining visual-language semantic understanding with action experts to generate continuous action sequences, and incorporates depth information to enhance fine manipulation. Reports indicate it supports various robot embodiments (including nine configurations), enabling cross-body transfer, and has released associated toolchains to lower training and reproduction barriers, positioning itself as a general-purpose 'brain' foundation.

Google Cloud has introduced 'parallel node pool auto-creation' for GKE, changing the previous sequential creation of multiple node pools into concurrent execution, thereby reducing provisioning latency for heterogeneous clusters and large-scale AI training/inference workloads. Internal benchmarks show up to an 85% reduction in overall configuration and scaling time for complex clusters, especially beneficial for workloads requiring diverse VM types or multi-host TPU slices that necessitate multiple independent node pools. This capability is integrated transparently into both Autopilot and Standard modes, and becomes effective upon upgrading to the specified GKE version.