Gemini CLI releases33Gemini CLI v0.67.0-nightly.20261022.g09d81bf11 released. The new version bridges the IDE companion to container sandboxes via exec stdio. Fixed 7 issues including repeated loading of agent when workspace is root directory, untrusted flag check scope, and IDE settings and GEMINI.md file being counted in context.
Hugging Face Daily Papers6060WildAWM enhances the generalization performance of general agents in web, CLI, and GUI benchmark tests through interactive synthesis of controllable computer environments and tasks. The system supports graphical interfaces, model context protocol services, and command-line interfaces, fully based on open-source models, and introduces an execution-based level mechanism to address solvability, difficulty maintenance, and verifier reliability issues in task synthesis. After training the Qwen model in the WildAWM synthetic environment, it achieved significant improvements in 8 cross-distribution benchmark tests.
Hugging Face Daily Papers7575PrismaForge enhances open-source models in visual encoding tasks to rival proprietary systems through scoring-based pipeline synthesis and filtering of high-quality multimodal code data; PrismaCoder-2M is currently the most diverse and largest open-source multimodal code dataset, with PrismaCoder-9B and PrismaCoder-35B outperforming the open-source field and matching proprietary systems after training on this dataset.
Browser Use3838Browser Use Cloud MCP server runs proxies in cloud browsers via API V3, offering 11 model selections, supports 10 parallel tasks, charges approximately 0.02 USD per minute, but only supports API V3 and has no GPT-6 Luna; its open-source version Browser Use open-source MCP runs locally for free, includes 16 direct control tools, allows custom models, suitable for long tasks.
Hugging Face Daily Papers6565SpatialScout evaluated the performance of VLM agents in active 3D object mapping, revealing significant flaws in their precision metric map construction. Claude Opus 5.5 scored an average overlap of 41.71% in this benchmark test, with 47.23% of target IoU below 0.5 and only 3.25% exceeding 0.9, showing a significant gap between recovering rough layouts and building complete, precise metric maps.
MarkTechPostRecursive releases the open-source 40B fuzzy model Flint, 2.5B active parameters, performs best among MoE models below 50B, supports inference and OpenAI compatible tool calls, 4-bit weights about 23GB, SWE-bench Verified 74.2%.
Codex releases33Rust 0.163.0-alpha.11.1 Released, Including Updates from 315 Commits
Claude Code releases1515Claude Code v2.1.307 is released, fixing over 20 bugs including API version migration, security event tracking, model alias error handling, prompt cache scope, MCP server authentication and timeout issues, and WebFetch tool error reporting, improving browser login, tool call retry, and self-hosted runner stability.
Ollama releases7070Recursive Flint v0.40.8 released, bringing code intelligence agents to Ollama. The 40B parameter MoE model supports 262K token context and achieves 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified. Flint can be used for code understanding, debugging, feature implementation, and testing, and supports MacOS.
AWS Machine Learning BlogThe Adaptive Agentic Intelligence Framework (AAIF) uses Strands Agents to process routine tasks with local small language models (SLMs), and routes complex requests to Amazon Bedrock cloud models, reducing costs while maintaining low latency.
TechCrunch · AISame story · 2 outlets7171Recursive releases its first model, Flint, a free coding model that can run on laptops, with almost the entire development process driven by AI. Recursive's automated research system, Eureka Machine, participated in almost every development step of Flint, enabling the team to complete the model construction in less than four months.
Sierra Blog7575𝜏-i18n expands 𝜏-Voice to five languages: Spanish, Brazilian Portuguese, Hindi, Korean, and Mandarin, evaluating cross-language performance using aviation, retail, and telecommunications tasks. The evaluation includes three dimensions: task completion, interaction effectiveness, and generation quality. The results show that Korean performed the weakest, GPT Live 1 had the highest task completion and generation quality, and Grok Voice Think Fast 2.0 had the best interaction but low generation quality.
PyTorch BlogTorchTitan adds native PyTorch support for the latest SOTA models including Muse Spark Open, Muse Glimmer, Qwen3-VL, Qwen3.5, Kimi K2.7, Kimi K3, and DeepSeek-V4.
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 BlogQwen3-4B's specialized vocabulary list VocabX reduces inference cost by 77% by focusing on the cost of each token rather than frequency, without changing the base model. This method for the Portuguese essay scorer reduces the output layer vocabulary from about 152,000 lines to about 65,000 lines, significantly reducing the computational overhead of token splitting and unused vocabulary items.
Baseten Blog5858Baseten Releases Global Top Decision Models mammoth-1 (27B) and elephant-1 (9B), Ranked 1st in Jev Decision Index 0.4 and 1st in 9B Decision Model respectively. Both are open-source and can be tried for free for one week on Baseten model API (default zero data retention).
OpenAI News5555Patch the Planet connects security researchers, maintainers, and incident responders to help address critical vulnerabilities in open-source projects.
Hugging Face BlogSame story · 2 outlets5858IBM released three models in the Granite Speech 5.0 series, with granite-speech-5.0-2b trained on 390,000 hours of data and supporting translation between 6 languages and 2.
How the accounts differ:Hugging Face Blog states the model is trained on 390,000 hours of data; MarkTechPost does not mention the training data duration.
Cohere Blog5858Cohere releases Rerank 5 Pro and Fast versions of its enterprise document retrieval models. The Pro version leads on major public benchmarks such as ViDoRe V3, nanoMTEB, BRIGHT, and BEIR, while the Fast version offers near-leading performance at a lower cost.
Ollama releases3535Ollama v0.40.8-rc0 has been released, adding support for the Flint model. It retains the source tool JSON and decoding API values to maintain numeric types and pattern order in prompts, while keeping the existing API serialization unchanged.
Ant Group inclusionAI · Hugging Face new models3535inclusionAI 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.
Latent SpaceAnthropic releases Claude Mods, allowing users to customize the behavior and interface of Claude Code through TypeScript or JavaScript functions, enabling personalized programming of attention interfaces.
NVIDIA Technical Blog3535The RFD4-Proteina can design octapeptide protein sequences, with over 2.5 billion combinations; 100-peptide sequences have possibilities exceeding the number of atoms in the observable universe, among which there are proteins that can be developed for new therapies; this model uses reward-guided reasoning to search for solutions to the protein design search problem.
Modal Blog7575Modal Dojo is a library for easily performing production-level LLM training, supporting models from 500 million to 10 trillion parameters, providing automatically optimized training recipes and built-in observability dashboards. Users can select models, load data, and start training through simple prompts, supporting data sources such as Hugging Face and Harbor environments, and can adjust recipes through parameters.
IBM Research BlogIBM releases vLLM-Hook MIA Plugin, implementing runtime monitoring and control in the vLLM inference engine, 20 times faster than its predecessor vLLM-Hook and 14 times faster than similar methods inserting code inside the model, with only a 2%-6% increase in latency. The plugin allows developers to observe and adjust model states in real-time, such as rejecting malicious prompts, marking hallucinatory facts, or editing response tone and accuracy, while almost not reducing performance.
NVIDIA BlogFeatured6262NVIDIA 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.
Why it matters: NVIDIA has rolled out open models across four lines: language, robotics, biology, and quantum, allowing readers to judge the coverage of its open-source ecosystem based on this.
NVIDIA BlogFeatured6262NVIDIA expands its open-source Agent Toolkit, adding components such as Agent Knowledge Foundation, the OpenShell secure runtime, and the NemoClaw blueprint, enabling enterprises to customize and govern agents with their own data.
Why it matters: The original article lists the new components of Agent Toolkit and details of deployments at multiple enterprises, showing how enterprise agents connect to their own data and are constrained by the runtime.
Mistral AI5050Mistral is collaborating with NVIDIA to integrate AI models with NVIDIA accelerated computing and physics modeling tools, accelerating engineering simulations and enabling closed-loop AI-driven design. The collaboration will launch public benchmarks and open datasets, standardizing the evaluation of physics and agent AI in computer-aided engineering, and leveraging NVIDIA Omniverse for 3D simulations and resource creation.
Hugging Face Blog5858NVIDIA launched SimReady Central Library on Hugging Face, aggregating over 2000 OpenUSD assets ready for direct simulation, covering robotic arms, grippers, warehousing equipment, racks, and cooling facilities, free to use without application.
NVIDIA Blog5858NVIDIA European partners are expanding regional computing capabilities to support model training and inference, helping developers build agentic and physical AI applications. Nscale achieves 37% more GPUs and 50% higher inference throughput using the NVIDIA DSX platform, Crusoe develops a liquid-cooled AI factory for low-latency inference, and Nebius adopts the NVIDIA Groq 3 LPX to accelerate sensitive applications.
NVIDIA BlogNVIDIA 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.
Databricks BlogDatabricks releases AI Runtime (AIR), bringing managed GPU infrastructure directly into its platform, allowing customers to train high-demand models at their enterprise data location. AIR provides GPU infrastructure, developer tools, and data governance features, supporting large-scale training, rapid iteration, and load safety, with customers not needing to manage a separate GPU stack.
NVIDIA Technical Blog6565NVIDIA OpenShell Reference Architecture ensures the security of AI agent workloads at scale by defining security boundaries and governance policies. The architecture allows agents to access files, tools, and the internet while performing complex tasks and enables dynamic collaboration. It mitigates risks associated with long-running agents through enhanced access control, audit logs, and permission management.
The Motubrain team from Shengshu Technology has launched a live stream of the 2.0 training process, lasting for half a month to showcase the training process of embodied models. The live stream covers four stages: pre-training, post-training, Agent task execution, and RSI exploration, opening the process from model training to capability iteration to the public. This move aims to enhance the transparency of Physical AI R&D and share research experiences in large-scale training and cross-task generalization.
AWS Machine Learning Blog3232Deepgram 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.
AWS Machine Learning Blog3131Since its launch, Amazon SageMaker JumpStart has integrated over 100 new models from 30 providers, including Qwen 3.8-27B, Cosmos3 series, etc.
AWS Machine Learning Blog3434The new AWS experience allows new SageMaker AI customers to enter a fully configured SageMaker Studio environment with just a few steps after registration, without setting up an advanced AWS account or manually configuring the environment. New customers also receive $100 in AWS Free Tier credit.
