Strata runs Qwen3.8-Flash-Next locally on consumer hardware
Niko1221 describes Strata as an open-source inference engine that runs the 125-billion-parameter Qwen3.8-Flash-Next model on consumer PCs. The project supports Windows and Linux, requiring a graphics card with at least 12 GB of VRAM. While 64 GB of RAM runs every model size, the developer notes that RAM determines which specific model fits. Strata provides a local API compatible with OpenAI and Anthropic formats, and includes optional image input. The developer measured performance on two gaming PCs, noting that an RTX 5070 achieves speeds exceeding 60 tokens per second. This setup allows local chat, code generation, and picture reading on standard hardware.
github.com
·
HN discussion
Researchers find most connected cars share data with third parties
A team from Northeastern University published the first large-scale measurement study on connected vehicle privacy. Working with Consumer Reports, they tested 21 late-model vehicles and 30 companion apps between October 2024 and August 2025. The study found that 19 of the 21 vehicles sent traffic to at least one third party, while seven apps transmitted sensitive identifiers to external companies. The authors note that once data leaves the device, consumers lose control over how it is shared or sold. This peer-reviewed work is scheduled for publication at IMC '26.
automatictransmission.khoury.northeastern.edu
·
HN discussion
Headstart accelerates Rust builds by starting dependent crates before dependencies finish type-checking
The Headstart project modifies rustc and cargo to begin compiling dependent crates immediately after a dependency's interface is checked, rather than waiting for full type-checking to complete. By writing an early metadata file, downstream crates can proceed with analysis while upstream function bodies are still being verified. This approach yields up to 54% faster cargo check and 42% faster cargo build times across several real-world projects. The implementation preserves existing error diagnostics and exit statuses, though it may increase memory usage and report errors slightly later. The patches are structured as a commit series intended for upstream pull requests.
github.com
·
HN discussion
Xray-core reporter alleges six-month cover-up of certificate verification flaw
A vulnerability reporter claims Xray-core concealed a certificate verification bypass for six months. The reporter states that after Xray-core removed the old pinning option in January 2026, users were forced to adopt a new option containing a flaw. Xray-core fixed the issue silently in February without disclosing the security risk. The reporter argues that this lack of transparency left users exposed to man-in-the-middle attacks until the issue was publicly reported in July 2026.
github.com
·
HN discussion
Google donates gVisor to CNCF
Google is donating the gVisor project, including its name and trademarks, to the Cloud Native Computing Foundation. The CNCF accepted the application on September 28, 2026. Over the next few weeks, gVisor will move to CNCF Sandbox status, migrate its build infrastructure to GitHub Actions, and add non-Google maintainers with merge permissions. This transition aims to prevent Google from having unilateral control over governance decisions and address adoption challenges related to corporate ownership.
gvisor.dev
·
HN discussion
JetBrains details Air Context RAG pipeline for semantic code search
JetBrains published the first installment of a series detailing the construction of Air Context, a retrieval-augmented generation pipeline designed to provide LLM agents with precise, citable evidence from code repositories. The post explains how the system parses and chunks source code using structure-aware methods for nine major languages, including Kotlin, Java, and Rust. It also describes the vectorization process, highlighting the trade-offs between dimensionality and precision in binary quantization. The authors note that Air Context is currently in public preview and included with JetBrains licenses, while future posts will address storage and query performance.
blog.jetbrains.com
·
HN discussion