Read this because This release addresses the critical data scarcity in Indian languages by providing a massive, auditable synthetic corpus and the reusable tools used to create it.
NVIDIA and AI4Bharat release a 2.8 trillion token Indic pretraining corpus and an open GPU-accelerated data generation pipeline for 22 languages.
Read this because This work addresses the critical gap in high-quality training data for multimodal coding, offering open models that rival proprietary systems through decomposed rubric synthesis.
PrismaCoder-9B and 35B models leverage a novel rubric-based data pipeline to lead open multimodal code intelligence benchmarks.
Read this because Holo4 offers a unified model for multi-interface automation, competing with frontier models at lower cost while open-sourcing training trajectories.
Hcompany introduces Holo4, a new series of agentic models supporting GUI, code, MCP, and API interfaces for generalist computer-use tasks.
Read this because Builders need this scale to train competitive multimodal models across video, audio, and image modalities without relying on proprietary data.
Researchers release LAION-BVD, a dataset containing 80 million videos totaling 10 million hours collected from CommonCrawl.
Cloudflare Infire (Rust) uses disaggregated prefill/decode to beat vLLM 0.10 by 20% on H100s. Unweight achieves 15–22% lossless model weight compression.
vLLM v0.20.0: 752 commits, 320 contributors. CUDA 13, PyTorch 2.11, Transformers v5, Python 3.14, FlashAttention 4 default, 2-bit KV cache.