#Model releases

inclusionAI/Ling-3.0-tiny-UBTree release: A fine-tuned model based on UBTree technology

inclusionAI released the UBTree fine-tuned version of Ling-3.0-tiny, Ling-3.0-tiny-UBTree-B16, with 183.1M parameters, combining a five-layer DFlash backbone and a bigram selector, achieving 2.6-3.5 times speed improvement on benchmarks like GSM8K and MATH-500. This model needs to be used in conjunction with the target model. The UBTree adaptation for the open-source version of SGLang has not yet been released.

NVIDIA expands its family of open models to cover language, robotics, biology, and quantum computing

NVIDIA expands its family of open models to cover language, robotics, biology, and quantum computing, including the new Nemotron 3.5 Super VL and Nemotron 3.5 Bolt.

NVIDIA Confidential Computing expands, providing secure extension capabilities for cutting-edge model builders across global infrastructure

NVIDIA Confidential Computing protects models and data execution through hardware root isolation and encrypted authentication, adding Confidential Bare Metal deployment options, enabling cutting-edge model builders to run sensitive workloads on third-party infrastructure while retaining model data control.

NVIDIA showcased its full lineup of local AI hardware: RTX Spark, DGX Spark, and DGX Station, helping to power next-generation intelligent agents.

NVIDIA showcased its full lineup of local AI hardware at GTC Berlin, with RTX Spark now available to support personal agents on Windows PCs, DGX Spark enabling developers to build and scale Linux applications, and DGX Station designed for Linux enterprise teams to run cutting-edge workloads, with the Windows version launching in December.

How Deepgram Optimizes Speech AI Performance with AWS Trainium2

Deepgram deploys the Nova-3 speech recognition model to PyTorch via AWS Trainium2, running 8 replicas on a single trn2.3xlarge instance, processing offline batch workloads with 11-70% more audio per dollar, and using the NKI library and Neuron Explorer/NAD toolchain for kernel-level tuning to improve inference efficiency.

i14 on how to use native PyTorch to train and deploy Laplace 3B linear attention model on AWS Trainium2.

i14 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.

The U.S. military's Cyber Command ECHOTRIBBLE project is turning AI agents into autonomous network weapons.

The U.S. military's Cyber Command (ARCYBER) is developing a platform named ECHOTRIBBLE, designed to transform AI agents into autonomous weapons capable of moving autonomously within other networks and 'neutralizing' targets. Internal documents obtained reveal that the project 'is designed for autonomous navigation of infrastructure' and can 'identify and neutralize identified threats/targets'. The system is expected to spread between devices like a computer worm until it finds the target and can operate on the cloud and 'networkless' isolated systems. Experts warn that such autonomous systems could cause uncontrollable collateral damage.

OpenAI claims that GPT-6 has significantly improved its performance on traditional programming languages such as COBOL and Fortran.

GPT-6 Astra achieved an average pass rate of 87.4% in MultiLangBench tests, an increase of 49.4% over GPT-5.6 Terra; GPT-6.1 Sol achieved an accuracy rate of 86.6% under low reasoning power and 97.9% under high reasoning power in the COBOL to C++ translation test.

Figma expanded MCP support but its founder considers it a non-perfect agreement.

Figma announced 20 new MCP partners at the Config inaugural conference in India, but its founder Dylan Field stated that MCP is not a perfect agreement and still needs improvement. Field expressed concerns about MCP's shortcomings in data sharing and noted that companies like Anthropic have not fully opened their resources. Figma believes that opening MCP may pose risks of abuse but remains committed to an open ecosystem.

Recursive Flint released on OpenHands: Open-source model helps local efficient coding workflows.

Recursive Flint is released, a hybrid expert model with 4 billion parameters, designed for efficient local inference, activating only about 250 million parameters per token, achieving 125+ tokens/second on M4 Pro MacBook Pro, and reaching 150+ tokens/second on NVIDIA RTX Spark. OpenHands, as its first partner, supports the model's operation on local hardware, providing developers with an open-source coding model option.

NVIDIA has released the Cosmos3-H-Dreams surgical generative simulation model.

NVIDIA has released Cosmos3-H-Dreams, extending interactive generative simulation to clinical gallbladder removal scenarios, capable of generating surgical scenes in real-time based on instrument movements.

Sundar Pichai: Google's mission boundaries, the scarcity of computing, and future AI products and businesses

Google CEO Sundar Pichai said that his latest released Gemini 4 Argon model has outperformed OpenAI and Anthropic's latest public models in coding, mathematics, science, and knowledge work benchmarks, bringing Google AI back to the top tier in the world.

DeepSeek V4.1 Flash 1017 released, US and China strengthen AI model regulation

CALIBER: Rethinking the Confidence of Reasoning Models

CALIBER estimates success probability before inference through training reasoning models and evaluates the confidence of specific answers after inference, significantly reducing calibration error. This 7B model on the BigMathDigits benchmark reduces ECE from 0.059 to 0.028 (a reduction of 52.5%), while achieving an AUROC of 0.806 and a Brier score of 0.130, proving that confidence evaluation is more reliable and can effectively distinguish between correct and incorrect answers.

Meta's TorchTitan integrates TorchTPU and optimizes TPU model training.

Meta's TorchTitan integrates TorchTPU, enabling PyTorch models to be trained natively on Google Cloud TPU with minimal code changes.

Meta's Prometheus Supercluster: The Journey to Build a Megawatt-Scale AI Infrastructure

HealthBench and HealthBench Professional: Evaluating AI model performance in real health conversations

OpenAI has released the HealthBench and HealthBench Professional evaluation benchmarks to assess the performance of AI models in real health conversations. In HealthBench Professional, GPT-6 Astra and GPT-6.1 Sol scored the highest after length adjustment (66.0% and 65.1%), while Claude Sonnet 5.5 scored the lowest (63.0%).

MiniMax M3.1 Flash released: Next-generation model with improved performance and reduced cost

MiniMax M3.1 Flash is officially launched, with nearly a threefold increase in online inference speed and a 70% reduction in cost compared to the previous generation. The model performs excellently in benchmark tests for complex intelligent tasks (cowork), computer use, software development, and network security, ranking highly in the AutomationBench-AA for automated tasks, scoring 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on ALE, and 82% on CyberGym.

Together Post-Training: Flexible and efficient model fine-tuning service

Together AI has launched Post-Training services, supporting the construction of custom models through reinforcement learning and supervised fine-tuning without complex infrastructure. The service provides flexible training loop definition capabilities and optimizes the underlying system and infrastructure to improve model quality and training speed. It supports mainstream open-source models such as Muse Spark Open, GLM 5.3, Kimi K3, and Qwen 3.8-27B, running on dedicated GPU resources to ensure efficient and numerically consistent training.