Model Releases

  • zai-org/GLM-5.3 — 🤖 MoE text-generation model with fp8 support, conversational capabilities in English and Chinese. Comes with eval results and endpoints compatibility.
  • tencent/Hy4-preview — 🤖 Hunyuan’s latest MoE preview for text generation and conversation. Apache-2.0 licensed, backed by two arxiv papers.
  • alibaba-pai/MiniMax-H3-Fun-Controlnet-Union — 🤖 Video generation model with ControlNet support for video-to-video, text-to-video, and image-text-to-video tasks. 🎬
  • BreezeBlue/Breeze-TTS-2 — 🤖 Text-to-speech model supporting voice cloning, design, and direction in English and Chinese. PyTorch/CUDA based. 🔊

Open Source Releases

  • tashfeenahmed/freellmapi — 🔥 One /v1 endpoint to rule them all: 34 free LLM providers, 635 endpoints, 7.4 billion tokens/month. Smart routing, automatic failover, encrypted keys. Built for personal experimentation.

Research Worth Reading

AI Dev Tools

  • abhigyanpatwari/GitNexus — 🛠️ Zero-server code intelligence engine running entirely in-browser. Drop in a git repo or ZIP file, get an interactive knowledge graph with built-in Graph RAG Agent. No servers, no setup, just code exploration. 🕸️

Today’s Synthesis

If you’ve been thinking about building domain-specific agents but don’t want to manage API keys for a dozen providers, freellmapi just solved the infrastructure problem — 34 free LLM endpoints behind a single /v1 with smart routing and failover. Pair that with the approach in CIFQA , which demonstrates how multi-agent pipelines with tool grounding and deterministic constraints can produce exact numerical answers instead of confident hallucinations, and you get a practical blueprint: use freellmapi’s routing layer to dispatch queries to specialized agent stacks depending on domain and complexity. Financial queries hit your CIFQA-style pipeline with formula verification; lighter tasks route to cheaper models. The failover matters too — when one provider drops, your agent pipeline doesn’t silently degrade. The takeaway for engineers: the hard part isn’t the model, it’s the routing and constraint logic around it, and both pieces are available today.