OpenSource, newest first.
OpenSource stories — newest first.
Microsoft Launches AI Behaviour Tester
Microsoft's open-source ASSERT framework automates application-specific AI behaviour testing, converting natural-language policies into scored test cases. This reduces manual overhead for platform engineers and security architects, ensuring AI systems adhere to defined constraints and organisational policies.
Luma AI launches open robotics lab
Luma AI's new open robotics lab offers engineers a shared platform to train robots, potentially reducing data acquisition costs and decentralising control over critical infrastructure. This challenges proprietary models and impacts supply chain resilience and national defence capabilities.
Releases tiny-vLLM engine on GitHub
Developer jmaczan released `tiny-vllm`, an open-source C++/CUDA LLM inference engine and course. It provides a practical guide to implementing high-performance techniques like PagedAttention and continuous batching, lowering the barrier for engineers to optimise LLM inference on GPU hardware.
Google releases Model Context Protocol server
Google's new open-source MCP server connects AI agents directly to Chrome Enterprise APIs, automating browser security management. This reduces manual effort for IT and security teams, streamlining tasks like DLP rule creation and security posture reviews, though human oversight remains critical.
whichllm releases tool to rank local LLMs
Selecting optimal local LLMs for deployment becomes more efficient. `whichllm` ranks models by real, recency-aware benchmarks, not just size, providing concrete data on VRAM fit, speed, and quality for platform engineers and procurement teams.
Release AI Networking Protocol MRC
OpenAI, AMD, and partners released Multipath Reliable Connection (MRC), an open protocol improving GPU networking for large AI training clusters. This standardises network reliability and efficiency, directly impacting training throughput and predictability for platform engineers.
DeepClaude enables cheaper AI models
Operational expenditure for AI-powered development drops significantly as `deepclaude` allows platform engineers and founders to leverage cheaper, high-performing models, cutting costs by up to 90% for Claude Code users.
trycua Open-Sources CUA Agent Framework
trycua's open-sourcing of CUA provides a unified, OS-agnostic framework for AI agents to control full desktops within sandboxed environments. This simplifies agent development and deployment across macOS, Linux, Windows, and Android, while raising new security considerations for desktop interaction.
Sachitrafa open-sources YourMemory agent layer
AI agents gain persistent, human-like memory through Sachitrafa's open-source YourMemory, a zero-infrastructure layer. It achieved 59% recall@5, outperforming Zep Cloud, and offers platform engineers a local, cost-effective solution for agent statefulness, reducing reliance on external memory services.
DeepSeek releases new V4 open-source AI model
DeepSeek's new V4 model claims rivalry with leading closed-source models, shifting AI development unit economics by introducing a new open-source alternative. This offers platform engineers and procurement teams a viable option, potentially impacting infrastructure expenditure and creating competitive pressure on global AI model pricing.
Intel archives open-source projects
Intel wound down its Open Ecosystem Community and Evangelism initiative, archiving numerous open-source GitHub repositories. This shift impacts platform engineers and developers, potentially increasing development timelines and integration costs for solutions built around Intel hardware.
Linux Kernel Cuts Legacy Network Drivers
Linux kernel developer Andrew Lunn proposed removing 27,646 lines of legacy network driver code due to an unsustainable surge in AI-generated bug reports. This shifts the kernel's maintenance philosophy, forcing developers to prioritise modern systems over old, largely unused hardware.
Qwen releases sparse coding MoE model
Advanced agentic coding and multimodal capabilities now require significantly less compute. Qwen's new sparse MoE model, Qwen3.6-35B-A3B, delivers performance competitive with larger models while enabling single-GPU deployment, reducing infrastructure costs for self-hosted AI.
China embraces, curbs OpenClaw AI
OpenClaw, an open-source AI agent, has seen rapid adoption in China, driving productivity and local tech giant engagement. However, government cybersecurity warnings and bans in state agencies now create a complex regulatory environment for AI adoption.
apfel unlocks Apple's Mac LLM
Apfel unlocks Apple's pre-installed on-device LLM on macOS 26, providing zero-cost, 100% on-device AI inference for Apple Silicon Macs. This offers platform engineers and developers a new privacy-preserving mechanism for local AI applications, bypassing cloud dependencies and API costs.
Miasma released to poison AI training data
Content creators gain a new defence against unauthorised AI data scraping with Miasma, an open-source tool that serves poisoned training data. This impacts founders and architects by offering a low-cost mechanism to degrade AI models, though it requires careful manual configuration.
Carmack Backs Open Source AI Training
John Carmack endorses AI training on open-source code, clarifying his stance as a prominent contributor. His view challenges anti-AI activist positions within the open-source community, framing code as foundational data for AI development.
Sarvam Open-Sources 30B, 105B Reasoning Models
Sarvam AI open-sourced its 30B and 105B MoE reasoning models, trained in India under the IndiaAI mission. This provides platform engineers with sovereign-trained, efficient models for diverse hardware, reducing inference costs and external reliance, particularly for Indian language applications.
New Git Extension Tracks AI Sessions
A new Git extension, `git-memento`, now records AI coding sessions as Git notes on commits. This establishes a critical mechanism for code provenance, directly impacting security architects and compliance officers by providing a verifiable audit trail for AI-assisted development.
Engineer Publishes Claude Code Workflow
A software engineer published a structured workflow for Anthropic’s Claude Code that strictly separates planning from execution. By forcing the AI to generate persistent, human-reviewed markdown files before writing code, developers maintain architectural control and prevent expensive system failures.
Sarvam releases five open-source models
CTOs and procurement teams gain high-performance, open-source alternatives to proprietary APIs. Local hosting reduces data residency risks and cuts inference costs because these models run on domestic infrastructure. This release validates significant regional investments, allowing Indian founders to build sovereign AI stacks without Western licence constraints.