
Google's gemma.cpp Runs Gemma 3 on CPU With No GPU Required
Google's gemma.cpp brings Gemma 2 and 3 inference to pure C++ with no GPU needed, targeting researchers who want to hack on model internals
20 stories

Google's gemma.cpp brings Gemma 2 and 3 inference to pure C++ with no GPU needed, targeting researchers who want to hack on model internals

NVIDIA's open-source NeMo Agent Toolkit adds inference-aware scheduling, framework-agnostic performance primitives, and a full build-evaluate-optimize loop for multi-agent systems

Kilo Code ran GLM-5.2 and Kimi K2.7 Code through the same backend task, and found the real gap is in planning, not building.

A local-first GGUF port turns OpenAI's Privacy Filter into a 1B PII detector that tags 55 categories in one pass, no Python required.

TaxHacker is a free, MIT-licensed, self-hosted app that uses LLMs to extract and organize data from receipts, invoices, and PDFs — with full control over your AI provider and prompts.

NVIDIA Container Toolkit v1.19.0 makes CDI the default GPU access method, adding IGX 2.0, Tegra CUDA Forward Compatibility, and read-only rootfs support

NVIDIA's official Kubernetes GPU device plugin hits v0.19.1 with WSL and CDI Tegra fixes, keeping GPU workloads schedulable at scale

Huobao Drama is an open-source TypeScript platform that turns a single sentence into a fully produced short drama, automating every step from script to final video using a pipeline of five AI agents.

NVIDIA GPU Operator v26.3.2 brings Kubernetes 1.36 support, richer GPU metrics with pod-level labels, and a new NRI plugin that simplifies GPU injection across non-standard Kubernetes distributions.

Tencent's open-source A.I.G platform scans AI infra, MCP servers, and agent workflows for CVEs, prompt injection, and jailbreak risks — all in one place.

Google open-sources Tunix, a JAX-native post-training library covering SFT, RL, distillation, and agentic training in one unified toolkit for TPUs

Anthropic's Claude Agent SDK gives Python developers a full agentic runtime with built-in tools, hooks, subagents, and in-process MCP servers — no framework assembly required.

Tencent ships ncnn 20260526 with Vulkan FlashAttention, broad BF16 coverage across all CPU backends, and first-ever HarmonyOS prebuilts

ByteDance's Pixel LLM codebase unifies SAM-2 and LLaVA into a single model that can segment, track, and converse about objects in images and videos — with a CVPR 2026 highlight paper on top.

Google's official MCP repo centralizes managed remote servers for BigQuery, GKE, Maps, and 10+ more services, with IAM auth and no local setup required.

The 186K-star project has quietly turned into a four-surface agent platform with chat-to-agent generation, a visual canvas, and a marketplace.

Baidu's PaddlePaddle 3.3 ships FlashMaskV3, a new checkpoint system, and dynamic GPU memory defragmentation for large-model training

Claude Fable 5 takes the top spot on DeepSWE's contamination-free, long-horizon coding benchmark, dethroning GPT-5.5 with a 70% pass@1 score on harder, more realistic tasks.

Anthropic's official claude-code-action brings Claude directly into GitHub PRs and issues, letting it review code, fix bugs, and open pull requests on command.

Chrome's HTML-in-Canvas API enters origin trial, letting live DOM elements render directly inside WebGL and WebGPU textures with full accessibility preserved