DeepSeek V4.1-Flash: Smarter, Faster, Cheaper — and It Beats Their Own Flagship
DeepSeek releases V4.1-Flash, a 552B asymmetric MoE that outperforms V4-Pro on benchmarks at lower cost, with a KV cache needing just 1/4 the HBM of the previous generation.
Hugging Face Releases @huggingface/kernels — 200+ WebGPU Kernels for Local AI in the Browser
Hugging Face introduces @huggingface/kernels, a library of 200+ WebGPU kernels enabling local AI inference directly in the browser without server-side compute.
Qwen3.8-27B: Frontier Coding Performance on a Single GPU
Alibaba's Qwen team releases an open-weight coding model achieving GPT-5.6 level performance on a single consumer GPU with 24GB VRAM.
Meta Returns to Open Weights with Muse Glimmer
Meta releases Muse Glimmer, a multimodal AI model designed to run on personal computers with 16GB+ RAM, marking a return to fully open weights.
MiniMax H3: One Open Model for Every Modality — 2K Video With Stereo Sound at a Third the Price
MiniMax launches H3, a general-purpose multimodal generation model that unifies text, image, video, and audio in one context — 15-second 2K video with native stereo sound, with open weights promised within days.
Kimi K3: Moonshot AI's Flagship Joins the Open-Weights Arena
Moonshot AI releases Kimi K3, its latest flagship model — landing as the largest open-weight model family competing with DeepSeek, Qwen, and GLM on agentic and reasoning benchmarks.
The Open Source Community Is Backing OpenEnv for Agentic RL
OpenEnv, the agentic RL environment library, is now coordinated by a committee including PyTorch, NVIDIA, Microsoft, and Hugging Face as a common protocol layer for RL environments.
Designing the hf CLI as an Agent-Optimized Way to Work With the Hub
Hugging Face rebuilt the hf CLI so the same commands serve humans and coding agents, cutting token use up to 6x on multi-step Hub tasks and trimming agent tool calls by roughly 30 percent.
Holo3.1: Fast & Local Computer Use Agents
Holo3.1 is H company's open computer-use model family with first quantized checkpoints for fast local agents, lifting AndroidWorld scores to 79.3% and cutting step times.
PaddleOCR 3.5: Running OCR and Document Parsing Tasks with a Transformers Backend
PaddleOCR 3.5 adds Hugging Face Transformers as a document parsing inference backend, letting PP-OCRv5 and PaddleOCR-VL 1.5 models run wherever the engine parameter is set.
Granite Embedding Multilingual R2: Open Apache 2.0 Multilingual Embeddings with 32K Context — Best Sub-100M Retrieval Quality
IBM Granite releases two Apache 2.0 multilingual embedding models on ModernBERT — a 97M model scoring 60.3 and a 311M model scoring 65.2 on MTEB Multilingual Retrieval.
Hermes Unlocks Self-Improving AI Agents, Powered by NVIDIA RTX PCs and DGX Spark
Hermes Agent from Nous Research crossed 140,000 GitHub stars in under three months, bringing self-improving AI agents to NVIDIA RTX PCs, workstations, and DGX Spark.