Open Source

PaddleOCR 3.5 introduces native Hugging Face Transformers integration, simplifying document parsing and layout analysis for PyTorch-based RAG workflows.

IBM open-sources its Granite 4.1 model family trained on 15T tokens, offering up to 512K context support under an Apache 2.0 license.

Microsoft has released VibeVoice-ASR-BitNet, compressing its speech recognition model to 1.58 GB for real-time CPU inference without dedicated GPUs.

IBM's open-source Granite Embedding Multilingual R2 model delivers a 32K context window and sets a new sub-100M MTEB retrieval benchmark score of 60.3.

A unix.foo developer post demonstrates how native iOS frameworks process text summaries locally, avoiding cloud API costs, latency, and privacy risks.

ServiceNow AI resolved a critical train-inference mismatch during a vLLM V1 engine migration, proving backend correctness must precede objective tweaks.

Hugging Face and TII have released QIMMA, a quality-first leaderboard with over 52,000 validated samples to accurately evaluate Arabic language LLMs.

An open-source repo called claude-token-efficient reports that custom CLAUDE.md files cut Claude API costs by up to 17% in its own self-reported benchmarks.

IBM has released Granite 4.0 3B Vision, an open-source 3-billion-parameter model designed to extract tables, key-value pairs, and chart data from enterprise files.

NVIDIA releases Nemotron 3 Nano 4B, a hybrid Mamba-Transformer model optimized for edge devices, offering up to 2x higher throughput and low VRAM usage.

ServiceNow and Hugging Face released EVA, an open-source framework that benchmarks 20 voice AI models across task accuracy and conversational experience.

NVIDIA NeMo Retriever achieves #1 on ViDoRe v3 with 69.22 NDCG@10 using an agentic loop and in-process thread-safe architectural optimizations.