🚀 RELEASES & LAUNCHES
Voodoo releases MIT-licensed dynamic quantization toolkit for LLMs
Voodoo released an MIT-licensed toolkit that learns per-tensor mixed-precision quantization for LLMs.- LLM developers and researchers can use Voodoo to create dynamic quantizations for models loadable by Transformers.
- The toolkit uses gradient descent to assign bit widths per tensor under a global size budget and exports standard GGUF files compatible with stock llama.cpp.
- The repository reports up to 89.5% lower perplexity and supports CUDA, ROCm, single-GPU, and tensor-parallel workflows.
Quixotic AI releases jinfer, an open-source JVM inference engine
Quixotic AI released jinfer, a pure-Java inference engine for multimodal AI models.- Java and JVM developers can run chat, vision, audio, embeddings, reranking, and text-to-speech models without Python or sidecar services.
- jinfer 0.2.0 supports GGUF and Safetensors, Spring AI and LangChain4j integrations, OpenAI-compatible serving, and model families including Gemma 4, Llama 3+, Qwen, and Liquid AI LFM 2.5.
- The stack uses Java Vector API or optional native kernels and supports GraalVM Native Image for self-contained binaries with millisecond startup.
- The release gives JVM builders an end-to-end local inference option implemented in Java, with quantizations including Q4_0 through Q8_0, MXFP4, F16, BF16, and F32.
⚠️ INCIDENTS & SAFETY
- CrofAI shuts down after report alleges silent routing to cheaper models
CrofAI shut down after a report alleged it routed requests to cheaper models.- The service targeted developers and other users seeking low-cost access to current AI models.
- An investigation reported that CrofAI wrapped OpenRouter and silently redirected requests to cheaper or weaker models, with markups of up to 20x.
- The reported substitutions included DeepSeek V4 Flash, GLM 5.3 Flash, Kimi K2.7 Code, Qwen3.7 Flash and Qwen3.5 9B.
- The owner acknowledged that a significant amount of traffic reached OpenRouter as a fallback before the service went offline, undermining its advertised model and pricing claims.