Ant Group inclusionAI · Hugging Face new models2222inclusionAI/ZwZ2-27B-GGUF model weights released, including language models and visual components with FP16, Q8_0, Q4_K_M precision, supporting llama.cpp multimodal inference.
Ant Group inclusionAI · Hugging Face new modelsThe inclusionAI/ZwZ2-4B-GGUF model weights have been released. This model is part of Zooming without Zooming v2 and is divided into a language model (FP16, Q8_0, Q4_K_M) and a visual encoder and projector (mmproj, FP16, Q8_0, Q4_K_M), offering multiple precision combinations. It needs to be used with llama.cpp (supporting Qwen3.5 multimodal).
Codex releases4545Codex 0.166.0 released, adding keyboard driver transcription selection (supporting word, line, and page operations), integrated search and configurable shortcuts, as well as consistent Markdown rendering in asynchronous questions, options, answers, and transcription responses. Also supports GPT-6.1 Terra and adds codex queue --resume feature to resume unloaded threads.
Hugging Face Blog5050Hugging Face releases open-source speech recognizer AnySpeechRecognizer and decision model AnyDecisionModel, supporting NVIDIA Parakeet and Alibaba Cloud Qwen3-ASR engines, and can run locally on Apple platforms.
ByteDance Seed · Hugging Face new modelsByteDance-Seed/GenFirst trains a latent VAE and combines it with a generative prior to avoid latent collapse, which usually makes end-to-end training unstable. The two-stage plan of this method is used for prior and entropy weights: the first stage lets the prior shape the latent geometry, and the second stage reduces both weights tenfold, allowing the decoder to recover details.
PyTorch Blog5858PyTorch® 2.15 released, defaulting to the rewritten NCCL backend, supporting fault logging and rollback; torch.distributed adds NIXL backend for unilateral transmission; FSDP2 and pipeline parallel optimize memory communication control.
Meta EngineeringMeta Andromeda's Next-Generation Ad Retrieval Engine Achieves Near-Ranking-Level Model Performance at the Retrieval Layer, Processing Over 1 Billion Ads Per Card Per Second, System Throughput Reaches 100 Billion Times Per Second, 100 Times Higher Than the 2024 Version, and Brings a 2.4% Conversion Rate Increase for Advertisers.
Ai2 BlogThe open-source model olmOCR 2 from Ai2 can be used to extract searchable, structured text from document images, including handwritten and complex layouts, and researchers have used it to make archive photo collections searchable and improve the transcription of handwritten archaeological records.
AWS Machine Learning Blog3434Exowatt customizes task-specific fine-tuned models Qwen3-1.7B and Qwen3-14B for edge and cloud deployment targets through Amazon SageMaker AI serverless models, achieving accuracy surpassing state-of-the-art models while reducing costs up to 49 times and shortening the model development cycle from months to days.
AWS Machine Learning Blog3434Property Finder discovered that 96.8% of the labels in its evaluation dataset for visual AI workloads migrated to Amazon Bedrock were predictions from a single model rather than human validation, hindering objective comparison of new models. The company built a trustworthy benchmark dataset by independently labeling 356 images with human experts and, after 27 iterative optimizations with evolutionary prompt tuning, reduced costs by 86% while maintaining an 86% accuracy rate.
Microsoft AI NewsFeatured6767Microsoft AI released the image editing model MAI Image 2.7, launching as a preview version. It ranks 3rd in Arena Edit and 2nd in the Artificial Analysis image editing chart, making Microsoft AI the 2nd place lab.
Why it matters: Microsoft announced the rankings of MAI Image 2.7 on two image editing lists and provided the improvement magnitude on 10 use cases compared to the previous generation.
LangChain Deep Agents releasesThe deepagents-code==0.2.2 version of LangChain Deep Agents has been released, including security and performance improvements. New features include GPT-6.1 Terra support, Chinese text input to chat, plugin update detection, etc.; bug fixes include project .env file trust value settings, terminal UI link limits, Git option limits, etc.; performance has been improved by reusing parent process context, optimizing server startup processes, and streamlining transcription rendering, etc. It is recommended to upgrade as soon as possible.
Apple Machine Learning ResearchThe GHARP method animates 3D heads in real-time using a small amount of input images and expression signals, divided into an offline identity phase (constructing geometry and appearance representations) and a runtime animation phase (predicting expression residuals). This method performs animation in the semantic structure latent space of the pre-trained reconstruction model, limiting the expression change space to achieve efficient residual prediction, and achieves SOTA quality on the Ava-256 benchmark, with a 13x speedup on A100 GPUs and an 8x reduction in the number of Gaussians.
PyTorch Blog3434Nemotron 3.5 achieved the highest streaming speech recognition accuracy on Google Pixel 11 using ExecuTorch and SME2, reaching up to 11.7% WER (with 1,120 ms audio blocks and QC8W precision), which is 3.5x faster than FP32 but with increased latency; shorter audio blocks (e.g., 80 ms) can reduce latency but accuracy drops to 15.2% WER.
Hugging Face BlogHugging Face releases torch-neuronx model supporting query Space, detecting AWS Trainium compatibility. This Space supports querying whether mainstream models (such as Qwen3-0.6B) with a scale of 175B parameters can run on Trainium without modification, and provides running scripts.
Tencent Hunyuan · Hugging Face new modelsTencent Mixuan released the WeVisOne-VL-4B model, built on Qwen3-VL, integrating a 9.6M parameter pixel decoder, supporting object-level perception (open-vocabulary object detection, pointing localization, counting, pointing and OCR) and pixel-level perception (pointing segmentation and promptable alpha friction), suitable for research tasks such as visual localization, pointing, counting, OCR localization, pointing segmentation, and image friction.
vLLM Blog5858vLLM provides zero-day support for MiniMax-M3.1-Flash on NVIDIA Blackwell, Rubin, and AMD MI350 series GPUs, performing excellently on agent workloads.
Hugging Face BlogHugging Face releases OlmoEarth v1.3, a satellite embedding model, offering global pre-computed embeddings (10m resolution, 64 and 128 dimensions), available to users without login or technical knowledge. The model performs best in 27 evaluations, with 128-dimensional embeddings scoring highest in 20 evaluations and 64-dimensional embeddings leading in 18. Training costs are reduced by 53% compared to v1.2 and 18% compared to v1, with corresponding reductions in carbon emissions. Weights, data, and code are all open source.
NVIDIA NewsroomEuropean transportation leaders Bolt with Lucid, Mercedes-Benz, Volkswagen Commercial Vehicles, and Einride are showcasing autonomous traffic applications on the NVIDIA Hyperion computing platform, demonstrating potential applications for Level 3 and Level 4 vehicles and future autonomous taxi fleets.
NVIDIA NewsroomAgile Robots has released the Agile ONE SN humanoid robot reference platform, equipped with NVIDIA Isaac GR00T and Jetson Thor, plans to integrate the NVIDIA Halos safety system, providing developers with a complete humanoid robot development and validation platform.
PyTorch Blog3030Torch-spyre's Inductor determines how each kernel's work is allocated on Spyre's core and manages the residency of tensors between on-chip Scratchpad and off-chip memory, and the backend compiler translates these decisions into instructions that hardware can actually run.
PyTorch BlogPyTorch users can accelerate dense and grouped GEMM, fused multi-head attention, MXFP8/MXFP4 scaled GEMM, RMSNorm, and TopK on AMD MI350 series GPUs using FlyDSL with existing operator API without calling custom kernels
Google ResearchGoogle Research's AI model is validated on JAMA Network Open for its accuracy in assessing gestational age using blind ultrasound scanning technology
Hugging Face BlogHugging Face introduced how to use TorchNeuron for system-level inference performance analysis on AWS Trainium, capturing clean traces with torch.profiler configured for the Neuron backend, and using key_averages() to identify time-consuming operations.
AWS Machine Learning Blog5858vLLM's bidirectional KV transmission mechanism reduces redundant calculations by allowing the decoding node to return previously computed KV cache to the pre-filled node in subsequent dialogue turns. In multi-round benchmark tests on AWS EC2 P5en instances, it improved the TTFT of Qwen3-32B by 1.6x and zai-org/GLM-5-FP8 by 3.3x. This feature is implemented by extending the NIXL KV connector and utilizing a KV cache-aware router to coordinate KV reuse across dialogue turns.
AWS Machine Learning Blog4949vLLM introduces a push-based NIXL connector, by moving post-pre-filled transmission coordination and setup out of the critical path, reducing time to the first token (TTFT) and time per output token (TPOT), enhancing application responsiveness.
AWS Machine Learning Blog3434Strategy Mosaic, as a semantic layer, models C360 domains on Amazon Redshift and Amazon Aurora without ETL using two linked models. Its metric syntax captures the business definition of agent reasoning and exposes the model to AgentCore-hosted Strands agents via the Mosaic MCP server, supporting user identity propagation to achieve end-to-end row-level and object-level security.
AWS Machine Learning Blog2828Denodo platform builds a semantic layer on AWS, providing controlled enterprise data access for Amazon Bedrock AgentCore, supporting federated queries across Amazon Aurora, Amazon Redshift, and local PostgreSQL databases.
AWS Machine Learning Blog3939i14 deploys its 24-layer decoder Laplace 3B (3,072D parameters, using Gated Linear Attention architecture) to AWS Trainium2, achieving a training speed of 120,000 tokens per second and concurrent inference capability for 11,544 output tokens.
vLLM BlogAnt Group and the vLLM community open-source Humming, a high-performance quantized inference kernel library. Humming supports NVIDIA GPUs starting from the Turing architecture and CUDA-compatible devices like the T-Head Zhenwu 810E, providing extensive quantization support covering various integer and floating-point types, bit widths, and scaling schemes. It significantly outperforms existing implementations such as vLLM Marlin, FlashInfer, and DeepGEMM on multiple representative workloads.
Apple Machine Learning ResearchTEAL has for the first time efficiently applied active learning to the token space of ViT for low-budget pixel-level semantic segmentation, boosting mIoU by +24.36 on datasets like CamVid and Cityscapes with only one pixel labeled per image, and achieving full-supervised performance of 97% with 10 pixels labeled.
Google Developers BlogArm leveraged the hierarchical graph visualization and plugin architecture of Google AI Edge Model Explorer to develop open-source adapters supporting formats such as TOSA, VGF, and PTE, helping developers more easily inspect edge model conversion processes.
Hugging Face BlogHugging Face has released TorchTPU, making TPU a native backend device for PyTorch. Transformers, Diffusers, and Kernels can now run directly on TPU. Users only need to select a model from the Hugging Face Hub and specify device_map="tpu".
Tencent Hunyuan · Hugging Face new modelsThe Hy-Memory-30B-A3B-GGUF model quantization file released by Tencent's Huayan uses a hybrid precision quantization scheme with imatrix calibration and MoE perception, achieving efficient local inference on llama.cpp, where the ~8 GB expert-mixed build file size is reduced to about 24.71 GB, with performance only 1.53% lower than Q6_K.
vLLM BlogvLLM achieves a downtime reduction of DeepSeek V4 Pro from 16GB200 to 20GB200 to 0.82 seconds and from 20GB200 back to 16GB200 to 1.96 seconds through asynchronous preparation and NIXL EP technology.
Google ResearchGoogle Research releases the VERID framework, enabling AI agents to run in physics-based, safe, and resource-efficient experiments; demonstrates closed-loop intelligent agent scientific discovery on adapter and de novo protein design prototypes; through three rounds of design-test-iterate cycles, AI-driven rational and de novo design will increase confirmed target hit rates to 33.5% and 42.1% on myoglobin and α-thrombin, respectively.
vLLM BlogvLLM has built a unified CuTe DSL decoding path for Kimi K3 through intelligent agent-assisted development, achieving a batch 1 performance of 840 tokens/s on 16 NVIDIA GB300 GPUs, with a 1.5-2x increase in decoding throughput in the test configuration, and a 1.8x increase in AgentX workload P90 interactivity.
Microsoft AI NewsMicrosoft AI has completed the pre-training of a new MoE large model MAI-Base-2, with 80 billion active parameters and 2.8 trillion total parameters, surpassing all open-source models in internal evaluations. The model is 4-12 times more efficient in pre-training FLOPs than the strongest open-source base model DeepSeek-V4-Pro-Base and is used to train larger models. Microsoft plans to license the model weights through Foundry using organic data pre-training and full-stack control.
Apple Machine Learning ResearchOn a cluster of four Mac Studio machines, Muse Spark Open (816B parameters) generates at a speed of over 36 tokens/second. The Mac Studio M5 Ultra version offers 4.6x faster inference speed (TTFT) and 1.6x faster tokens/second compared to the M3 Ultra, benefiting from the GPU neural accelerator and 1.2 TB/s memory bandwidth.
