HuggingFace's Top Trends of the Week - What's Open Sourced
TL;DR.
New: Ripple-65B trained from scratch and strong in the terminal. Arete-27B weaned Qwen from government framing. TwIL-8B Math Olympiad. The P-H3-Image drawing is based on the MiniMax video model. Viggle Remake recreates the finished video. Kimi K3.1 Pro has purchased quanta.
מחזיק: Idol, LTX-2.5, Qwen3.8-27B, FLUX 3 Klein, Nemotron 3.5 Transcribe, Qwen3.5-4B-3D-Coder.
Masters and agents.
Ripple-65B-Preview - a dense model with 65 billion parameters without MoE, which American Recursive trained from scratch for 12 trillion tokens; I mentioned this in the OpenRouter review. Weights under Apache 2.0 weigh 129.6 GB in BF16 and 57.2 GB in NVFP4, the context of up to a million tokens. According to the authors' measurements in Terminal-Bench 2.1, it has 80.5, the best result among the American open models in their comparison, but the Chinese could not get there. You can compete on OpenRouter for $0.20 and $0.60 per million tokens.
Arete-27B. Qwen3.8-27B, which Nos Research has additionally qualified to answer political issues without government framing and refusals. According to the authors' test, the share of answers with Chinese country framing dropped from 22% to 0.9%, and in general tests the model remained at the baseline level. Users noticed a side effect: more often Arete bases confidently invent things about things that do not exist. Apache 2.0, 54.7 GB ב-bf16.
TwIL-8B is a math reasoning model and Olympiad code based on IBM Granite 4.2-8B. According to the authors' measurements on MATH-500, it has 98.2, and GGUF in Q4_K_M weighs 5.3 GB. In learning tests it is inferior to larger samples, only declared English, the license is not commercial.
Kimi K3.1 Pro, which I disassembled after the scales were released, acquired within a week Quanta on GGUF, EXL3 and MLX.
Photos and videos.
P-H3-Image creates and edits images in 8 steps, and is made of the MiniMax H3 video model; You can send up to eight photos for editing.