Jianzhi Robotics open-sourced RealOmni, the largest no-embodiment dataset in the embodied AI industry, covering 10 major scenarios and over 30 skills, containing more than 10,000 hours and over 1 million high-quality multimodal data entries; the first batch of data has been launched on the Baidu Baihe AI computing platform, supporting one-click distributed training startup; Baidu Baihe became the first cloud platform to fully support the three major open-source embodied VLA models: RDT, π0, and GR00T N1.5, providing engineering optimizations such as 35.0%+ training acceleration for π0.5 and 36%+ inference acceleration for world models.
NVIDIA released Cosmos Reason 2, an open-source reasoning vision-language model (VLM) for physical AI, which ranks first on the Physical AI Bench and Physical Reasoning leaderboards.
TII and Hugging Face jointly released the Falcon-H1-Arabic model family (3B/7B/34B), which features a pioneering hybrid Mamba-Transformer architecture and context windows of up to 128K–256K tokens. It significantly enhances long-text coherence, dialect coverage (including Egyptian, Levantine, Gulf, etc.), and STEM reasoning capabilities, setting new SOTA benchmarks across Arabic evaluation suites such as OALL, 3LM, ArabCulture, and AraDice.
At CES 2026, NVIDIA showcased a desktop-level embodied agent solution based on DGX Spark and Reachy Mini. The system utilizes the Nemotron 3 Nano reasoning model, the Nemotron Nano 2 VL vision model, and ElevenLabs TTS, implementing multi-model routing and tool calling via the NeMo Agent Toolkit. It also includes fully open-source code and a step-by-step deployment guide.
In the article, Arvid Kahl argues that software products should enhance user trust and retention by simplifying data import and export (such as CSV, Apache/Nginx configuration exports), rather than creating vendor lock-in. He cites his own SaaS practices, including Permanent Link (a URL shortening and archiving service), Fathom Analytics data migration, Podscan, and Feedback Panda, to illustrate that open data rights lower the barrier to user decision-making, increase initial adoption rates, support user-driven analysis and disaster recovery, and ultimately foster more sustainable customer relationships.
SGLang v0.5.7 has been officially released, adding Day 0 support for models including Mimo-V2-Flash, Nemotron-Nano-v3, LLaDA 2.0, the Qwen-Image-Edit series, and Qwen-Image-Layered.
Tongyi Lab has launched Qwen-Image-2512, the December update of the Qwen-Image text-to-image base model. Compared to the initial version released in August, it features improvements in human texture (such as facial details, wrinkles, hair strands) and natural textures (such as water flow.
At the Create 2025 conference, Baidu Intelligent Cloud launched the Kunlun Chip Super Node, integrating 64 Kunlun XPUs into a single chassis. Inter-card connectivity bandwidth increased by 8 times, training performance per chassis improved by 10 times, and inference performance per card improved by 13 times. It supports a 1U 4-card liquid cooling design, enabling maintenance of a single node by one person, reducing XPU temperature by over 20°C, and can be deployed in air-cooled data centers using the Tianji 1.0 CDU.
SGLang officially released Model Gateway v0.3.0, featuring a refactored six-layer metrics architecture, UUID-based worker resource management, support for all router types in the Unified Inference Gateway (IGW) mode, new tokenize/detokenize HTTP interfaces, a GLM-4 function calling parser, native embeddings support, server-side TLS, a Go implementation version, and other major enhancements, along with a migration checklist.
ServiceNow releases AprielGuard, an 8B parameter security guard model designed to detect 16 categories of security risks (such as toxicity, illegal activities, privacy violations, etc.) and various adversarial attacks (such as prompt injection, jailbreaking, memory hijacking, multi-agent exploitation sequences). It supports three input formats: standalone prompts, multi-turn conversations, and agent workflows (including tool calls and reasoning traces), and offers a reasoning mode (interpretable classification) and a fast mode (low-latency classification).
Baidu's new model ERNIE-5.0-Preview-1203 topped the LMArena text leaderboard as the number one in China with a score of 1451, surpassing several mainstream models including Claude-Opus-4-1 and GPT-5.2. The model has 2.4 trillion parameters and is a native omni-modal large model supporting input and output for text, images, audio, and video. Baidu Qianfan simultaneously enhanced its enterprise-level Agent Infrastructure, providing model services, an MCP marketplace, a data intelligence platform, and operational assurance capabilities.
The Tongyi team released Qwen-Image-Edit-2511, which improves upon Qwen-Image-Edit-2509 by enhancing character consistency (including multi-person group photos), integrating community LoRA capabilities (such as lighting adjustments and new viewpoint generation), and strengthening industrial design and geometric reasoning abilities. The image editing feature is available online via Qwen Chat, while optimal local performance requires deployment through ModelScope. The cited technical report is arXiv:2508.02324.
The Tongyi Qianwen team has released two new models in the Qwen3-TTS family: the voice creation model Qwen3-TTS-VD-Flash and the voice cloning model Qwen3-TTS-VC-Flash, both available via the Qwen API.
As an independent developer, the author shares six machine-assisted marketing methods practiced in the Podscan project: programmatic SEO (publishing podcast transcript pages and guiding sign-ups), pattern exposure (generating industry trend charts based on business data), AI similar-customer mining (reverse-engineering target communities and content scenarios from existing customers), AI customer scoring (automatically scoring and filtering high-intent customers based on user behavior), building SEO-friendly documentation (knowledge base serving as both a resource and acquisition channel), and competitor channel reuse (tracking how B2B competitors reach audiences on platforms like LinkedIn).
QwenTeam has launched Qwen-Image-Layered, a multimodal model capable of decomposing a single image into multiple RGBA layers. It supports a variable number of layers and recursive decomposition; each layer can undergo high-fidelity editing operations such as recoloring, object replacement, text modification, deletion, scaling, and moving independently. The paper has been published on arXiv:2512.15603.
vLLM released version v0.13.0, including 442 commits and 207 contributors (61 new contributors); added support for models such as BAGEL, AudioFlamingo3, and JAIS 2.
Hugging Face's official blog announced a major refactoring of the tokenizer system in Transformers v5, with the core change being the separation of the tokenizer architecture (normalizer, pre-tokenizer.
Mistral AI has released Mistral OCR 3, achieving an overall win rate of 74% compared to Mistral OCR 2. It features significant enhancements in tables, handwriting, scanned documents, and forms, supporting Markdown+HTML table reconstruction.
NVIDIA released the Nemotron 3 Nano 30B A3B model and simultaneously open-sourced the NeMo Evaluator library along with complete YAML configurations, logs, and structured artifacts.
On December 17, 2025, Google DeepMind released Gemini 3 Flash, a new speed-optimized model in the Gemini 3 series that combines Pro-level reasoning capabilities with Flash-level latency/cost efficiency.
Developers can now submit apps to ChatGPT for review and publication; approved apps will appear in a newly created app directory within the product. OpenAI has simultaneously updated its tools, guidelines, and Apps SDK to support building chat-native experiences that trigger real-world actions.
Google DeepMind has released Gemma Scope 2, an open-source interpretability toolkit for the entire Gemma 3 series (from 270M to 27B parameters). It supports Sparse Autoencoders (SAEs) and cross-layer transcoders, enabling analysis of safety-related behaviors such as hallucinations, jailbreaks, refusal mechanisms, and chain-of-thought faithfulness. The suite was trained on approximately 110 PB of data with over 1 trillion total parameters; interactive demos are provided by Neuronpedia.
Google Research developed and experimentally deployed the Paper Assistant Tool (PAT) for STOC 2026, a specialized AI tool based on Gemini 2.5 Deep Think.
IBM Research's open-source configurable general agent, CUGA, has officially launched on Hugging Face Spaces. It supports open models such as gpt-oss-120b, integrates Langflow visual orchestration, MCP/LangChain multi-tool calling, and computer use capabilities, achieving leading results in the AppWorld and WebArena benchmarks.