1. Stanford and Lambda Labs Release OpenJarvis Local Agent Framework
Researchers at Stanford University and Lambda Labs have released OpenJarvis, an open-source, local-first framework for running inference, agents, memory, and learning entirely on-device. The framework uses a declarative configuration object called a "spec" to decompose AI systems into five swappable primitives: Intelligence, Engine, Agents, Tools & Memory, and Learning. OpenJarvis optimizes local specs using a frontier cloud model as a teacher during search time, resulting in zero cloud calls during inference. The framework supports 11 local models across four families, including Qwen3.5, Gemma4, Nemotron, and Granite.
- • Released under the Apache 2.0 license on March 12, 2026.
- • Performs within 3.2 percentage points of the best cloud models with 800x lower marginal API cost and 4x lower latency.
- • Uses a declarative configuration spec to decompose AI systems into five swappable primitives.
- • Supports 11 local models across Qwen3.5, Gemma4, Nemotron, and Granite families.
- • Provides built-in support for 25+ data connectors and 32+ messaging channels.
It allows developers to build local-first personal AI agents that perform within 3.2 percentage points of top cloud models while reducing API costs by 800x.
2. Huawei Open-Sources KVarN for 3-5x KV-Cache Compression in vLLM
Huawei has open-sourced KVarN, a KV-cache quantization method designed for agentic and long-context workloads. Implemented as a native vLLM attention backend, KVarN integrates into the vLLM framework using a single flag and requires no model changes, retraining, or calibration. The system quantizes the KV cache using a four-stage process, with a default configuration (kvarn_k4v2_g128) that uses 4-bit keys and 2-bit values. KVarN delivers 3-5x more context capacity and up to 1.3x the throughput of FP16 while maintaining FP16-level output quality and reasoning accuracy.
- • Released under the Apache 2.0 license and built on vLLM v0.22.0.
- • Provides 3-5x more context capacity compared to the ~2x capacity offered by FP8.
- • Achieves up to ~1.4x FP16 throughput while maintaining FP16-level output quality.
- • Delivers up to ~2.4x the throughput of TurboQuant while maintaining higher accuracy on reasoning tasks.
- • Integrates into vLLM using a single flag with no model changes or calibration required.
It allows developers to run longer-context and agentic workloads on vLLM with significantly reduced memory footprints and up to 1.3x higher throughput than FP16.
3. Boxes.dev Launches Cloud-Only Environments for Claude Code and Codex
Founders Nick and Drew have launched boxes.dev, a cloud-only agentic development environment (ADE) that provides dedicated remote compute for running Codex and Claude Code agents. The platform allows developers to run resource-intensive coding agents on remote compute using snapshots of their full development environment, bypassing local hardware constraints and git worktree management issues. Boxes.dev features a desktop app, a mobile app, scheduled automations, and a Slack integration.
- • Provides dedicated cloud computers for Codex and Claude Code agents.
- • Runs agents on remote compute using snapshots of the full development environment.
- • Features a desktop app, mobile app, scheduled automations, and Slack integration.
- • Developed by the former co-founder/CTO and first hire of Gem.
It allows developers to run resource-intensive coding agents in parallel without cluttering local git worktrees or hitting local hardware constraints.
4. Anthropic Open-Sources Reference Implementation for AI-Powered Vulnerability Discovery
Anthropic has released an unmaintained reference implementation for autonomous vulnerability discovery and remediation using Claude. The repository provides a multi-stage pipeline (Build, Recon, Find, Verify, Dedupe, Report, and Patch) that uses gVisor sandboxing to isolate autonomous agents during execution. The pipeline is specifically configured for finding C/C++ memory vulnerabilities using Docker and ASAN, and supports Claude APIs across Bedrock, Vertex, and Azure.
- • Provides a reference implementation for autonomous recon, find, triage, report, and patch processes.
- • Uses gVisor sandboxing to isolate autonomous agents during execution.
- • Supports Claude APIs including Bedrock, Vertex, and Azure.
- • Specifically configured for finding C/C++ memory vulnerabilities using Docker and ASAN.
- • The repository is unmaintained and does not accept contributions.
It provides developers with a concrete, multi-stage pipeline architecture to build their own autonomous security and patching agents.
5. Miso Labs Releases MisoTTS 8B Open-Weights Emotive Speech Model
Miso Labs has released MisoTTS, an 8-billion-parameter open-weights text-to-speech model under a modified MIT license. The model utilizes a residual vector quantization (RVQ) architecture composed of a 7.7B backbone for temporal prediction and a 300M decoder for depth prediction. By conditioning on both text and audio context, MisoTTS can respond dynamically to the tone of a speaker. Miso Labs claims a latency of 110ms, compared to 300ms for Sesame and 700ms for ElevenLabs, though the model is currently limited to half-duplex, single-turn interactions.
- • Released under a modified MIT license with open weights.
- • Features 8 billion parameters (7.7B backbone, 300M decoder).
- • Utilizes a residual vector quantization (RVQ) architecture with 32 audio codebooks.
- • Conditions on both text and audio context to respond to speaker tone.
- • Claims a latency of 110ms, compared to 300ms for Sesame and 700ms for ElevenLabs.
It enables developers to build highly responsive voice agents, claiming a latency of 110ms compared to ElevenLabs' 700ms.
6. NVIDIA Releases LocateAnything 3B for Local UI and Object Grounding
NVIDIA has released LocateAnything 3B, a lightweight model designed to run locally for UI and object grounding. The model combines grounding, OCR, and UI understanding to instantly locate objects, buttons, or text based on natural language verbal descriptions. It is optimized for local deployment to support fast, on-device automation and interface interaction.
- • 3B parameter model designed to run locally.
- • Combines grounding, OCR, and UI understanding.
- • Instantly locates objects, buttons, or text based on verbal descriptions.
It enables developers to build fast, local agentic workflows that can interact with user interfaces and documents without relying on cloud APIs.
7. BeeLlama v0.3.1 Integrates Upstream llama.cpp Features and DFlash Speedups
BeeLlama versions 0.3.0 and 0.3.1 have been released with architectural updates to align with upstream llama.cpp. The update integrates features such as Multi-Token Prediction (MTP) and Gemma 4 12B support, while improving DFlash to handle multi-slot and multi-GPU configurations. The release also adds support for q6_0 KV cache and TQ3_1S and TQ4_1S model quantization options, providing prebuilt binaries and Docker images for all major platforms.
- • Aligns with upstream llama.cpp and integrates Multi-Token Prediction (MTP).
- • Adds support for Gemma 4 12B and q6_0 KV cache.
- • Improves DFlash to handle multi-slot and multi-GPU configurations.
- • Achieves up to 4.93x speedups for Qwen 3.6 27B and Gemma 4 31B on a single RTX 3090.
- • Provides prebuilt binaries and Docker images for all major platforms.
Developers running local models can leverage up to 4.93x speedups for Qwen 3.6 27B and Gemma 4 31B on consumer hardware like a single RTX 3090.
8. Anthropic Details Security Containment and Sandboxing for Claude Code
Anthropic has detailed its security practices for containing agentic products like Claude Code and Claude Cowork. To mitigate risks such as user misuse, model misbehavior, and external attacks, Claude Code utilizes OS-level sandboxing (Seatbelt on macOS and bubblewrap on Linux), which has reduced permission prompts by 84%. Anthropic also disclosed past vulnerabilities where project-local configuration was parsed before establishing a trust boundary, and demonstrated via an internal red-team exercise that an employee could be phished into exfiltrating AWS credentials via Claude Code.
- • Claude Code utilizes OS-level sandboxing (Seatbelt on macOS, bubblewrap on Linux) to reduce permission prompts by 84%.
- • Claude Code auto mode is designed to catch approximately 83% of overeager agent behaviors before execution.
- • Disclosed past vulnerabilities where project-local configuration was parsed before establishing a user trust boundary.
- • Internal red-team exercise demonstrated that an employee could be phished into exfiltrating AWS credentials via Claude Code.
- • Claude Cowork employs a full virtual machine architecture using Apple's Virtualization framework on macOS or HCS on Windows.
It provides critical security context for developers running Claude Code locally, highlighting past vulnerabilities in project-local configurations and the importance of environment-layer containment.
9. Gradio 6.16.0 Released with Security Patches and MCP Updates
Gradio version 6.16.0 has been released, introducing a configurable heartbeat feature via the GRADIO_HEARTBEAT_INTERVAL environment variable and updating the MCP endpoint to display a browser landing page. The release implements critical security patches for path traversal in gr.FileExplorer, an open-redirect bypass in OAuth, and SSRF in Image, Gallery, and Audio post-processing. It also includes bug fixes for Dataframe and Tabs browser freezes.
- • Patches path traversal in gr.FileExplorer and open-redirect bypass in OAuth.
- • Patches SSRF in Image, Gallery, and Audio post-processing.
- • Introduces configurable heartbeat via the GRADIO_HEARTBEAT_INTERVAL environment variable.
- • Updates the MCP endpoint to display a landing page when visited via a browser.
- • Fixes browser freezes in Dataframe and Tabs.
Developers using Gradio should update to patch vulnerabilities including path traversal in FileExplorer and SSRF in image/audio post-processing.
10. NVIDIA Releases Agentic Safety Dataset for Indirect Prompt Injection Testing
NVIDIA has released an agentic safety dataset on Hugging Face consisting of 1,272 synthetic red-teaming records. The dataset covers nine distinct enterprise domains and is designed to test the ability of tool-using agents to resist indirect prompt injections hidden within tool-returned data.
- • Consists of 1,272 synthetic red-teaming records.
- • Covers nine distinct enterprise domains.
- • Designed to test resistance to indirect prompt injections hidden in tool-returned data.
- • Published on the Hugging Face platform.
It provides developers with a concrete dataset to evaluate and harden their tool-using agents against indirect prompt injections hidden in retrieved data.
11. Higgs Audio v3 TTS 4B Released for Multilingual Voice Chat
Higgs Audio v3 TTS 4B is a text-to-speech model built specifically for voice chat applications. The model supports 100 languages and includes support for inline control, allowing developers to build highly interactive and multilingual conversational audio interfaces.
- • 4B parameter text-to-speech model.
- • Built specifically for voice chat applications.
- • Supports 100 languages.
- • Includes support for inline control.
It provides developers with a lightweight, highly multilingual model optimized for real-time conversational audio interfaces.
12. Alibaba Releases Qwen-Image-Flash for Fast Image Generation
Alibaba has released Qwen-Image-Flash, a model designed for fast, high-quality text-to-image generation and instruction-guided editing. The model utilizes few-step distillation, data composition, teacher guidance, and task mixture to achieve high performance and low latency.
- • Designed for fast, high-quality text-to-image generation.
- • Supports instruction-guided image editing.
- • Utilizes few-step distillation, data composition, teacher guidance, and task mixture.
It provides a lightweight, fast option for developers integrating image generation and editing capabilities into their applications.
13. SynthTraces Launches to Generate Synthetic Coding Agent Session Traces
SynthTraces is a new minimal codebase designed to generate synthetic coding agent session traces using Pi. The project uses a harness where an open model acts as a coding agent with read and bash access to open-source Hugging Face projects, while a small local model running on llama.cpp acts as a human user to prompt the agent. The project has generated and published over 2,000 session traces on Hugging Face for training or fine-tuning LLMs.
- • Minimal codebase for generating synthetic coding agent session traces.
- • Uses an open model as a coding agent with read and bash access.
- • Uses a local model running on llama.cpp to act as a human user.
- • Generated over 2,000 Pi session traces published on Hugging Face.
It provides developers with a dataset of over 2,000 session traces on Hugging Face to train or fine-tune their own task-specific LLMs and coding agents.