Hugging Face AI Updates: September 4, 2026

1. Hugging Face Will Be Acquired by NVIDIA for $12.9 Billion

Hugging Face. NVIDIA confirmed a roughly $12.93 billion deal to acquire the model hub, ending weeks of reports. NVIDIA committed to keeping the platform independent, framework-neutral, and open to all clouds and hardware accelerators, with its branding and mission intact and no requirement that developers deploy on NVIDIA compute. Jensen Huang framed the investment around platform reliability, safety, and model evaluation infrastructure for the 18 million developers and 200,000 companies currently on the hub. Source

2. Funes Gives Coding Agents a Searchable Memory of Past Sessions

Hugging Face. Funes is a single-binary memory layer for coding agents including Claude Code, Codex, pi, and Hermes. It indexes session traces locally using embeddings plus BM25, reranks with a cross-encoder, and exposes recall and get tools after running funes add [agent]. Memories stay local or sync to a private Hugging Face dataset for cross-machine access, with credentials redacted before publishing. In testing, recall came out 8x cheaper than a written handoff on one task and 4x on another, and it preserves original evidence rather than a summarized version. Source

3. NeoMME Drops the Vision Tower for a Single Multimodal Transformer

H Company. NeoMME is a family of 260M and 800M parameter encoders where one bidirectional Transformer processes text tokens and raw 32x32 image patches together, with no separate pretrained vision tower. It supports dynamic image resolution, a 16,384-token context, alternating sliding-window and global-attention layers, and a 131k-token multilingual vocabulary trained from scratch. The 260M variant hits 0.523 nDCG@10 on ViDoRe v3, beating every model under 800M parameters, and encodes about 51 pages per second at 2048x2048 on an L40S. Both sizes ship Apache 2.0 with Transformers support. Source

4. 100 GRPO Steps Move a 350M Model From 22.6% to 29.7% on IFStruct

Hugging Face. A walkthrough fine-tunes LiquidAI’s LFM2.5-350M for structured output generation using GRPO via TRL, training 100 steps on roughly 500 samples with LoRA adapters covering about 6M parameters, or 1.66% of the model. Three weighted reward functions score JSON format validity, field count accuracy, and schema compliance, with training data drawn from NVIDIA’s Nemotron-RL structured-outputs set. Overall IFStruct accuracy rises from 22.6% to 29.7%, with JSON formatting jumping from 18.0% to 31.9%. Source

5. A GRPO Pipeline Trains Qwen3.5-35B to Write Watercolour Code

Hugging Face. Sergio Paniego published an open reproduction of a viral watercolour-painting project, training Qwen3.5-35B-A3B with LoRA and GRPO to emit JavaScript that paints using the p5.brush library. The composite reward combines a compile-and-render gate, a sketch-length term, a pairwise judge scored against a hand-curated pool of 178 reference paintings, and the HPSv3 aesthetic preference model. Three runs with different reward weightings all improved, with the judge-led run gaining +0.27 reward over 110 steps. Code, models, datasets, and training infrastructure are published openly. Source