Model Releases

  • LiquidAI LFM2.5-24B-A2B MoE Language Model — LiquidAI’s LFM2.5-24B-A2B is a Mixture-of-Experts language model optimized for edge deployment, supporting multilingual conversational AI and text generation, with architecture detailed in arXiv 2511.23404.
  • MiniCPM-V 4.7-1B On-Device Multimodal Model — MiniCPM-V 4.7-1B is a compact, on-device multimodal model supporting image-text conversation, designed for lightweight deployment on edge devices while maintaining conversational capabilities through transformers and safetensors.
  • Fishaudio Drama-3 Multilingual TTS Model — Fishaudio Drama-3 is a multilingual text-to-speech and audio-editing model supporting emotional speech generation, multi-speaker scenarios, and integrates bsgemm-moe optimization for efficient inference, with arXiv 2610.18010 detailing the architecture.

Open Source Releases

Today’s Synthesis

When inheriting an ML codebase shaped by previous LLM assistance, you can systematically audit logic, flag risky code, and structure docs for retrieval by running three targeted tools in sequence. astscribe 0.12.0 analyzes notebook structure, variable flows, and cell dependencies without adding another LLM layer, giving you a clean map of what each cell actually does. Immediately follow with doc-ai-slopcheck 1.2.0 , which applies 73 deterministic rules across Python, JavaScript/TypeScript, Go, and C/C++ to surface common AI-generated antipatterns—missing imports, unsafe assumptions, dead logic—before they reach production. Finally, graph-knowledge-doc-parser 0.2.16 ingests your READMEs and design docs, using graph awareness to resolve entities, preserve structure, and feed clean context into your RAG pipeline. The engineer-actionable idea: configure these as a post-commit CI step. You’ll get a triage report stating exactly what the notebook accomplishes, where the generated code is fragile, and which documentation chunks are queryable, turning opaque AI-assisted workflows into auditable, recoverable artifacts.