HuggingFace's Top Trends of the Week - Coming to Opensource New TL;DR: Flint to code agents for 2.5 billion actives. Nanbeige 4.3-33B with 512,000 Spark-X2.6-4B in the context of agents on her device. Smart-A1.5 for research purposes. MiniMAX M3.1 Flash temporary weights from Mistral for time series. Indra OCR on the processor. Chromos on top of a MOSS image. Compatible: DeepSeek-V4.1-Flash-1017 and its quantum lines, FLUX 3 Klein, MOSS-Image, LTX-2.5, Qwen-Image-2.1, Qwen3.8-27B, Drama 3. LLM and Flint Agents - Agent coding model, designed for one graphics card or Mac: MoE, 2.5 billion active out of 40, 262,000 tokens, Apache 2.0. The verified SWE-bench measurements are 74.2%. llama.cpp and vLLM supported it in two days, the Unsloth Q4 quantum weighs 25.1 GB, and OpenRouter gives it away for free. . Praise speed, expose the contrived claims of tools without a rigid conversation plan and thinking of 20,000 tokens before the code. Nanbeige 4.3-33B is a 33 billion dollar compressed model of the recruitment team at BOSS Zhipin. In the authors' table in Terminal-Bench 2.1, it has 80.9 compared to 73.0 for Qwen3.8-27B. Section 512 thousand, apache 2.0, but without an expensive mustache: the discussions counted 26 GB in 262 thousand tokens on q8 compared to 8.6 GB Qwen. Model Spark-X2.6-4B for 4 billion agents in its device: the context 1 million, three levels of thinking, Q4 quantum weighs 3.5 GB. According to the authors' measurements, twice as big as Qwen3.5-9B in agent tests it is bypassed, and the infinite logic of the previous version, according to reviews, has disappeared. Works in llama.cpp and Ollama, you can feel in the demo. Sage-A1.5 is an agent for long-term research missions from the AI ​​laboratory in Shanghai, based on Qwen3.6-35B-A3B-A3B: it divides parts of its work into sections and checks the conclusions. In GAIA the authors state 95.1. The official GGUF: Q4 weighs 21.7 GB and fits a 24 GB card. MiniMAX M3.1 Flash, which broke last week, received the promised weight: MoE, 23 billion active out of 428, 252 GB.

HuggingFace's Top Trends of the Week - Coming to Opensource New TL;DR: Flint to code agents for 2.5 billion actives. Nanbeige 4.3-33B with 512,000 Spark-X2.6-4B in the context of agents on her device. Smart-A1.5 for research purposes. MiniMAX M3.1 Flash temporary weights from Mistral for time series. Indra OCR on the processor. Chromos on top of a MOSS image. Compatible: DeepSeek-V4.1-Flash-1017 and its quantum lines, FLUX 3 Klein, MOSS-Image, LTX-2.5, Qwen-Image-2.1, Qwen3.8-27B, Drama 3. LLM and Flint Agents - Agent coding model, designed for one graphics card or Mac: MoE, 2.5 billion active out of 40, 262,000 tokens, Apache 2.0. The verified SWE-bench measurements are 74.2%. llama.cpp and vLLM supported it in two days, the Unsloth Q4 quantum weighs 25.1 GB, and OpenRouter gives it away for free. . Praise speed, expose the contrived claims of tools without a rigid conversation plan and thinking of 20,000 tokens before the code. Nanbeige 4.3-33B is a 33 billion dollar compressed model of the recruitment team at BOSS Zhipin. In the authors' table in Terminal-Bench 2.1, it has 80.9 compared to 73.0 for Qwen3.8-27B. Section 512 thousand, apache 2.0, but without an expensive mustache: the discussions counted 26 GB in 262 thousand tokens on q8 compared to 8.6 GB Qwen. Model Spark-X2.6-4B for 4 billion agents in its device: the context 1 million, three levels of thinking, Q4 quantum weighs 3.5 GB. According to the authors' measurements, twice as big as Qwen3.5-9B in agent tests it is bypassed, and the infinite logic of the previous version, according to reviews, has disappeared. Works in llama.cpp and Ollama, you can feel in the demo. Sage-A1.5 is an agent for long-term research missions from the AI ​​laboratory in Shanghai, based on Qwen3.6-35B-A3B-A3B: it divides parts of its work into sections and checks the conclusions. In GAIA the authors state 95.1. The official GGUF: Q4 weighs 21.7 GB and fits a 24 GB card. MiniMAX M3.1 Flash, which broke last week, received the promised weight: MoE, 23 billion active out of 428, 252 GB.