SmolDocling vs VibeVoice

SmolDocling is a 256-million-parameter vision-language model from IBM Granite that converts documents — PDFs, scanned papers, tables, charts, forms — into clean, structured text with remarkable accuracy for its size. It introduces a new markup format called DocTags that captures not just text but document structure, reading order, and element types (headings, captions, tables, code blocks) in a way that downstream models and parsers can reliably consume. The "smol" in the name is intentional: at 256M parameters, SmolDocling runs fast enough to be deployed in production pipelines where larger VLMs would be prohibitively slow or expensive. Despite its compact size, IBM reports it achieves state-of-the-art performance across multiple document type benchmarks — outperforming much larger models on structured document parsing tasks. The key innovation is the DocTags format, which gives the model a precise vocabulary for describing document elements rather than trying to reconstruct structure from freeform text output. Built on top of the docling project (58.7k GitHub stars), SmolDocling is open source under Apache 2.0 and available on HuggingFace. The technical report is on arXiv (2503.11576). For teams building RAG pipelines, document intelligence tools, or any system that needs to ingest unstructured documents at scale, this is a practical, deployable solution.

VibeVoice is Microsoft's open-source family of frontier voice AI models covering both speech recognition and synthesis at a scale most commercial services still can't match. The ASR model processes up to 60 minutes of audio in a single pass, generating speaker-diarized, timestamped transcriptions across 50+ languages — complete with hotword customization for domain-specific accuracy. At 7B parameters, it supports on-premise deployment for privacy-sensitive applications. The TTS side is equally impressive: VibeVoice-1.5B synthesizes up to 90 minutes of multi-speaker audio with natural conversational flow and turn-taking between up to four distinct speakers. A lightweight 500M realtime variant streams at under 300ms latency. All of this runs on a novel continuous speech tokenizer operating at just 7.5 Hz — dramatically more efficient than typical audio codecs. What makes this notable is the MIT license. Microsoft isn't just open-sourcing a research demo; they're releasing production-grade weights on Hugging Face alongside code that teams can self-host, fine-tune, or build into their products. With 42,000+ GitHub stars and 771 earned today alone, it's the kind of drop that resets the baseline for what open-source audio AI looks like.