NAll news

All news

Monday7 items

  1. China.com TechnologyModel releases
    AI score77

    Meituan Releases LongCat-2.5 Large Model with 1.6 Trillion Total Parameters

    Meituan has released its new-generation large model LongCat-2.5, a natively multimodal model designed for long-horizon agent tasks, with 1.6 trillion total parameters, trained entirely on domestic compute. Evaluation results show that the model's scores on multiple mainstream benchmarks fall between Claude-Opus-4.8 and Claude-Opus-5, and it surpasses Claude-Opus-5 on some evaluations. Meituan has also launched an AI business operations workbench called "Miaoshou" based on the LongCat series, helping local-life merchants understand industry metrics and brand rules, and call tools within authorized scope to complete business tasks.

    Same story · 4 reports

    #Model releases#Multimodal#Agents

  2. Phoenix TechnologyModel releases
    AI score55

    Du Xiaoman and Hong Kong University of Science and Technology Jointly Release Xuanyuan 4.5 Financial-Enhanced Large Model

    Du Xiaoman, the Hong Kong University of Science and Technology, and the Hong Kong Generative AI Research and Development Center (HKGAI) jointly released the Xuanyuan 4.5 financial-enhanced large model at the opening ceremony of the 2026 Hong Kong FinTech Week. Based on RSI technology, it improves effectiveness by 30% in core financial scenarios such as customer acquisition, risk control, and customer management, while costing only 10% of a general-purpose model. The model consists of two parts: a large language model and a multimodal model. The self-developed Xuanyuan-Machine participates in agent data synthesis, supervised fine-tuning, reinforcement learning, and other stages, and has built more than 100 billion tokens of scenario-enhanced data. Based on Xuanyuan, Du Xiaoman has built over a thousand agents and more than 5,000 skills, with average daily token usage exceeding 100 billion, and has shortened the iteration cycle of risk strategies in its credit business by more than half.

    #Model releases#Agents#English text is being prepared.

  3. Sina TechModel releases
    AI score68

    Meituan Releases LongCat-2.5 Large Model, 1.6 Trillion Parameters Trained Entirely on Domestic Compute

    Meituan has released its next-generation large model LongCat-2.5, a natively multimodal model designed for long-horizon agent tasks, with 1.6 trillion total parameters, trained entirely on domestic compute. During the mid-training stage, the model conducted long-context training with sequences of up to 1M tokens, with cumulative training volume reaching trillions of tokens, and it optimized task division around supernodes to address the challenge of small per-card memory in domestic compute. Meituan says its results on multiple mainstream benchmarks fall between Claude-Opus-4.8 and Claude-Opus-5, and it surpasses Claude-Opus-5 on some evaluations.

    Same story · 4 reports

    #Model releases#Multimodal#Agents

  4. English text is being prepared.Model releases
    AI score68

    Meituan LongCat releases LongCat-2.5 native multimodal model, 1.6T parameters targeting long-horizon agent tasks

    Meituan LongCat releases the native multimodal model LongCat-2.5, with a total parameter count of 1.6 trillion and approximately 48 billion parameters activated per token, strengthening native visual understanding and agentic coding capabilities on the basis of LongCat-2.0. The model incorporates more long-horizon trajectory data into long-sequence training, and through large-scale long-horizon reinforcement learning enhances end-to-end engineering capabilities such as complex codebase development and complete project construction, while representing visual signals as multi-level discrete tokens to uniformly process visual and textual inputs. The official statement says it has optimized long-context training and W4A4 inference on domestic compute, and it will soon be available on GitHub and HuggingFace.

    Same story · 4 reports

    #Model releases#Multimodal#Agents

  5. TMTPost · AGIModel releases
    AI score56

    DeepSeek V4.1 Flash-1017 Released with Public Pareto Frontier Curve

    DeepSeek launched a new version of its large model, V4.1 Flash-1017, and for the first time independently published a performance-versus-cost Pareto frontier curve, claiming that the comprehensive cost per unit task is only one-hundredth to one-twentieth that of comparable models such as Fable and Astra. In the DeepSWE v1.1 evaluation, the new version's highest pass rate Pass@1 reached 77.0%, compared with about 74% for the previous version, with the highest score occurring at an average cost of about 1.3 yuan per task. The public API pricing remains unchanged this time; the cost benefits come from token efficiency optimizations such as task-aware compression, dynamic context trimming, and adaptive decoding, with actual measurements showing token usage in most scenarios dropping by more than 50% compared with the previous version. The article also points out that the curve was drawn by the vendor itself, and that the task set, the evaluation framework DeepSeek Harness, and the inference-tier rules are all internally defined, so whether cross-model comparisons can be reproduced remains to be verified.

    #Model releases#DeepSeek#Reasoning

  6. Phoenix TechnologyModel releases
    AI score67

    Meituan LongCat releases the LongCat-2.5 native multimodal model, with 1.6T parameters targeting long-horizon agent tasks

    Meituan LongCat has released its native multimodal model LongCat-2.5, with a total parameter count of 1.6 trillion and approximately 48 billion parameters activated per token, strengthening native visual understanding and agentic coding capabilities on the basis of LongCat-2.0. The model incorporates more long-horizon trajectory data into long-sequence training and, through large-scale long-horizon reinforcement learning, improves end-to-end engineering capabilities such as complex codebase development and complete project construction. LongCat-2.5 represents visual signals as multi-level discrete tokens, enabling the backbone model to process visual and textual inputs in a unified token form and supporting tasks such as creative design and 3D.

    Same story · 4 reports

    #Model releases#Multimodal#Agents

  7. Testing CatalogModel releases
    AI score57

    Nace AI Open-Sources Multimodal Decision Model Drex-Max v1, Runs Locally on M4 and M5 Macs

    Nace AI has open-sourced the multimodal decision model Drex-Max v1, which can run entirely locally on Apple Silicon M4 and M5 Macs, while also offering a cloud inference entry point. To use it, provide the model with a document and a set of questions with fixed answers—for example, which team should handle a given support ticket—and the model returns a probability for each answer.

    Quoted postNace AI · @NaceAI

    We’re open-sourcing Drex-Max v1 The most capable Multimodal Decision Model ! And yes: it runs 100% locally on Apple Silicon (M4 & M5) Enjoy Building with Nace! Cloud Inference: https://t.co/fm3GXfjaqx Local Inference: https://t.co/9nwcBn7c2D https://t.co/Wgu8P4zByR

    #Model releases#Open-source ecosystem#English text is being prepared.

Sunday3 items

  1. LeiphoneModel releases
    AI score60

    Pika Open-Sources H3 Flex Based on MiniMax H3, Achieving Joint Audio-Video Generation in Six Steps

    Pika has released Pika H3 Flex, a derivative model of MiniMax H3, opening the adapter weights and inference code, and jointly generating visuals and sound through six-step denoising. The model includes two sets of LoRA adapters, natively supports 480P and 768P output, can generate approximately 5 to 15 seconds of video with synchronized stereo audio, and provides a ComfyUI workflow and a Hugging Face Diffusers integration method. The team first used supervised fine-tuning, then used Projected Distribution Matching Distillation (PDMD) to compress 50-step denoising to 6 steps; high-resolution enhancement and audio track separation are provided by the Pika API.

    #Model releases#Open-source ecosystem#Video

  2. LeiphoneModel releases
    AI score60

    SenseTime Open-Sources SenseNova Loop-9B-Preview Looped Computation Model

    SenseTime has open-sourced SenseNova Loop-9B-Preview, which adopts the Looped Transformer architecture, retains the original decoder modules of the first and last 4 layers each, and replaces the middle layers with parameter-sharing looped computation modules, expanding the effective computational depth through multiple rounds of iteration and benchmarking against 27B models at the 9B parameter scale. The model introduces mHC (Manifold-Constrained Hyper-Connections) multi-path residual connections to enhance feature interaction; in ablation experiments, GPQA Diamond and TeamBench improved by 4.42 points and 13.34 points, respectively. SenseTime claims that the model outperforms other models in code generation, office scenarios, and agent tasks, and surpasses Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench. The model weights have been open-sourced.

    Same story · 3 reports

    #Model releases#Open-source ecosystem#Reasoning

  3. English text is being prepared.Model releases
    AI score60

    SenseTime Open-Sources SenseNova Loop-9B-Preview, Claiming It Rivals 27B-Parameter Models

    SenseTime announced on October 30 the open-sourcing of SenseNova Loop-9B-Preview. The model adopts a Looped Transformer architecture and expands effective computational depth through recurrent computation; the company claims that at only 9B parameters it can rival 27B-parameter models. The model retains the original decoder modules of 4 layers each at the beginning and end, converts the middle layers into parameter-sharing recurrent computation modules, and introduces mHC (Manifold-Constrained Hyper-Connections) to enhance feature interaction. The company claims that it surpasses Qwen3.5-27B on evaluations such as WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench, and that in ablation experiments GPQA Diamond and TeamBench improved by 4.42 points and 13.34 points, respectively.

    Same story · 3 reports

    #Model releases#Open-source ecosystem#Agents

Saturday5 items

  1. Testing CatalogModel releases
    AI score65

    Rock Matrix Open-Sources EdgeRock-4B Text-Image Reasoning Model

    Rock Matrix has open-sourced EdgeRock-4B, a small text and image reasoning model with fixed-capacity memory. The model fully retains the most recent 1,024 tokens and stores earlier important information in a separate memory bank, thereby keeping the attention KV cache size fixed and preventing long conversations and long reasoning from continuously consuming more GPU memory. Rock Matrix reports an average pass@1 of 82.9% on AIME 2024–2026, compared with 18.8% for the strongest bounded-state model it tested.

    #Model releases#Open-source ecosystem#Reasoning

  2. Phoenix TechnologyModel releases
    AI score63

    Xiaomi Open-Sources 0.8B Document Understanding Model Xiaomi-OCR-0

    Xiaomi has released and open-sourced a lightweight document understanding model, Xiaomi-OCR-0, with only 0.8B parameters, integrating three capabilities—document parsing, key information extraction, and OCR visual question answering—into one. The model scored 95.24 on Real5-OmniDocBench, ranking first among compared models, and scored 96.83 on OmniDocBench v1.6, ranking first among models under 1B parameters, with an average score of 83.2 across five mainstream OCR visual question answering benchmarks. Xiaomi has also open-sourced the model weights and inference code, and provides a Demo, MCP service, and Agent Skill toolkit. The related technology has already been used in Xiaomi phones and "Super Xiao Ai" for screen information recognition such as pickup codes, bills, and cards.

    Same story · 2 reports

    #Model releases#Open-source ecosystem#Multimodal

  3. Aravind SrinivasModel releases
    AI score55

    Perplexity CEO Shares Early Evaluation of Gemini 4 Carbon: WANDR Benchmark Sets New SoTA

    Perplexity CEO Aravind Srinivas shares an early evaluation of Gemini 4 Carbon, calling it the strongest model to date in long-form research and answer quality, and noting that it achieves a new SoTA on the Wide-And-Deep-Research (WANDR) benchmark. The model will be made available to all Perplexity Pro and Max users once token capacity is sufficient, integrated into Computer.

    #Evaluations / benchmarks#Reasoning#Google

  4. SiliconANGLEModel releases
    AI score55

    Krafton's Ludo Robotics Releases Robot Intelligence Model Ludi 0.2

    Krafton subsidiary Ludo Robotics has released the robot AI system Ludi 0.2, which builds on Ludi 0.1 launched in August with enhanced social intelligence, memory and reasoning, tool coordination, gesture interpretation, and physical mobility capabilities. The system combines the reasoning and dialogue capabilities of vision-language models with navigation and manipulation abilities, with the goal of enabling robots to understand human intent, respond to feedback, and adapt to changing situations. The company says it is advancing toward a final version that integrates perception, dialogue, and physical actions into a single "brain."

    #Model releases#Embodied AI#Multimodal

  5. OpenAI: Deployment Safety and System CardFirst-partyModel releases
    AI score72

    OpenAI releases the GPT-Rosalind-6 system card addendum, offering restricted access for biological research

    OpenAI releases the GPT-Rosalind-6 system card addendum. This model uses the same set of weights as GPT-6 Astra, with no new training data or post-training added, and is provided only through trusted access to vetted research institutions, enterprises, and government partners. At the highest reasoning budget, MedChemBench v7 averages 49.6% (GPT-Rosalind-5.5: 28.9%), GeneBench Pro v13 single-attempt success rate is 37.7% (previous generation: 16.7%), and LifeSciBench Gold average score is 61.3% (previous generation: 44.0%). Under the Preparedness Framework, this release is rated High capability in the biological and chemical domains, below the Critical threshold; rated Critical in the cybersecurity domain; and below High in AI self-improvement.

    #Safety / alignment#OpenAI#English text is being prepared.

Friday77 items

  1. English text is being prepared.Model releases
    AI score72

    Xiaomi open-sources the 0.8B-parameter OCR model Xiaomi-OCR-0, supporting Markdown and JSON output

    Xiaomi has open-sourced the lightweight document understanding model Xiaomi-OCR-0, with a parameter scale of 0.8B; its weights and inference code are now available on Hugging Face and GitHub. The model is built on Qwen3.5-0.8B-Base and trained on approximately 170 million OCR and document understanding samples. A single model covers three types of tasks: document parsing, key information extraction, and OCR visual question answering, and it natively supports Markdown, LaTeX, and JSON output. Official evaluations show that it scores 96.83 on OmniDocBench v1.6, ranking second on the overall leaderboard and first among models with fewer than 1B parameters. It also provides a browser Demo, an MCP service, and an Agent Skill toolkit.

    Same story · 2 reports

    #Model releases#Open-source ecosystem#Multimodal

  2. SenseTimeModel releases
    AI score63

    SenseTime Open-Sources the SenseNova Loop-9B-Preview Model

    SenseTime has open-sourced SenseNova Loop-9B-Preview, a 9B Looped Transformer that improves information flow and reasoning capabilities through parameter-sharing loops and mHC (Manifold-Constrained Hyper-Connections). The company claims it achieves 27B-level performance on complex tasks, surpassing Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench; mHC improves the GPQA Diamond score by 4.42 points and TeamBench by 13.34 points. The model is now available on Hugging Face and GitHub, with plans to later expand to larger models and adaptive fast/slow thinking.

    Same story · 3 reports

    #Model releases#Open-source ecosystem#Reasoning

  3. Cailian PressModel releases
    AI score73

    Mistral Releases Large 4 Open-Weight Model, Focused on Coding and Cyber Defense

    French company Mistral AI has released the Large 4 model (ML4), featuring advanced coding and cyber defense capabilities; enterprises can download the open-weight version and deploy it to run on their own infrastructure. The model has a total of 1.05 trillion parameters, with approximately 52 billion parameters actually activated per task execution, and was trained in a European data center using 3,800 NVIDIA Grace Blackwell GPUs. Mistral claims it is comparable to other leading open models on coding tasks and emphasizes that open source helps prevent frontier AI technology from concentrating in the hands of a few players.

    Same story · 15 reports

    #Model releases#Open-source ecosystem#Agents

  4. Shanghai ObserverModel releases
    AI score74

    Zhiyuan Releases Embodied Model GO-3 Preview, Plans Full Open Source by End of 2026

    Zhiyuan has released its new-generation robot embodied model GO-3 Preview, which connects the full pipeline from training and simulation practice to continuous learning after deployment, ranking first in three tests: RoboDojo, RoboTwin, and RoboColiseum, with plans for full open source by the end of 2026. In real convenience store scenario tests, the Elf G2 robot's task success rate was 28.6% on the first day and rose to 85.7% after five days; for the small hook hanging task, the success rate was 17.7 percentage points higher than with traditional training methods.

    Same story · 12 reports

    #Model releases#Embodied AI#Open-source ecosystem

  5. Shanghai ObserverModel releases
    AI score64

    Daxiao Robotics and Nanyang Technological University S-Lab Release World Model Mundo

    Daxiao Robotics, in collaboration with Nanyang Technological University S-Lab, has released Mundo, a world model built for specific environments that can expand 1 to 12 images into explorable, interactive three-dimensional scenes. The model writes physical properties such as object mass and friction into the scene program, with a physics engine computing motion and collisions, and it supports editing objects, properties, and actions with code. Mundo has been used to reconstruct hotel floors and laundry areas, providing simulation environments for robot navigation and manipulation training.

    Same story · 4 reports

    #Model releases#Embodied AI#Multimodal