KTransformers v0.7.0 shipped with broader hardware compatibility and new fine-tuning capabilities for large Mixture-of-Experts models. The release is the most notable open-source move today.
The ecosystem gained 238,868 GitHub stars across the 10,000+ tracked projects so far today. 146 new projects entered the index. This week 2,363 new projects and 100 new Hugging Face spaces were added.
The top gainer on GitHub today is speech-to-speech with plus 6,339 stars.
KTransformers v0.7.0 brings expanded MoE fine-tuning and deployment options. The update improves hardware compatibility for running large sparse models. llama.cpp b10456 fixes thread and block counts in quantized kernel launches. The adjustment reduces under or over allocation. Morphic v1.6.0 resolves citation issues when the model prepends a toolu prefix.
KTransformers v0.7.0 featureKTransformers v0.7.0: Expanded MoE Fine-Tuning and Deployment · KTransformers v0.7.0 brings broader hardware compatibility and new fine-tuning capabilities for large MoE...
Morphic v1.6.0 What's Changed * fix: resolve citations when model prepends a toolu_ prefix by @miurla in https://github.com/miurla/morphic/pull/888 * fix: handle compound TLDs in citation...
llama.cpp b10456 sycl: fix thread/block count in quantized cpy kernel launches (#27160) · Adjusts the thread/block count to be proportional to the size of the quant, reducing under/over...
Alibaba cut Qwen3.6 27B input price by 50 percent to 0.30 dollars per 1M tokens. Qwen3.6 35B A3B input dropped 7 percent to 0.14 dollars. GLM 5.2 from Z.AI fell 33 percent to 0.31 dollars. Nemotron 3.5 Lightning from NVIDIA fell 20 percent to 0.08 dollars. Alibaba raised Qwen2.5 VL 72B Instruct input price by 220 percent to 0.80 dollars. Qwen3 30B A3B input rose 8 percent to 0.13 dollars. DeepSeek V4 Pro 0813 input rose 50 percent to 0.66 dollars.
cc-haha leads new detections with 14,136 stars. It is a local-first cross-platform desktop workspace for Claude Code and agents with multi-agent support and a skill marketplace. Observal reached 2,361 stars as a local registry and analytics platform for AI components. tokenspeed hit 1,908 stars as a speed-of-light LLM inference engine. LetsFG reached 1,739 stars for agent-native flight and hotel search. Citadel reached 903 stars as an operating layer for Claude Code and OpenAI Codex.
LetsFGPython · 1,739 stars · Agent-native flight & hotel search and booking — MCP server, CLI, and Python/JS SDKs. Hundreds of airlines...
CitadelJavaScript · 903 stars · The operating layer for Claude Code + OpenAI Codex: persistent project memory, intent routing, safety...
tokenspeedPython · 1,908 stars · TokenSpeed is a speed-of-light LLM inference engine.
ObservalPython · 2,361 stars · Observal is a local registry and analytics platform for your AI components. Setup Observal, define the...
cc-hahaTypeScript · 14,136 stars · Local-first cross-platform desktop workspace for Claude Code / agents: multi-agent, Git worktrees, code...
A Hugging Face community paper with 52 upvotes evaluates defenses against AI-generated video attacks on real-world crisis events. Recent video generators can fabricate realistic depictions of wars and disasters. Existing benchmarks provide limited evidence on detector performance.
Today's facts show a market splitting into two directions: MoE fine-tuning tools mature while inference prices diverge sharply across Alibaba's model lineup. The surge in agent infrastructure projects signals growing demand for operational layers around coding agents.
Source: olud.ai tracking of 10,000+ open-source AI projects, 300+ models and live provider pricing. Figures are measured, not estimated. All releases · Live pricing · Latest in AI
← All AI news