xClean Tools

AI Models — Daily Top 5

UPDATED 2026-11-02 01:15 PST

2026-10-26 · MONDAY · 13:16 PDT

  1. 1 MemTensor/MemT-1.0-Nano 30,338 DOWNLOADS · 142 LIKES MemTensor published MemT-1.0-Nano, the small entry in its MemT-1.0 line of memory-native language models. Built on Qwen3.6-35B-A3B and licensed Apache 2.0, it ships a separate memory block weight file and custom modelling code alongside the main shards, and is documented in an accompanying technical report. The repository has drawn 30,338 downloads and 142 likes.
  2. 2 WarmBloodAban/Minimax-h3_Singularity 783,234 DOWNLOADS · 915 LIKES Singularity is a community fine-tune of MiniMax-H3 for reference-to-video generation, distributed as BF16, INT8, pruned INT8 and w4a8 checkpoints for ComfyUI. The V2 update widens action coverage, tightens prompt adherence and fixes audio reference conditioning, and the repository ships its own prompt-writing specification. It has drawn 783,234 downloads and 915 likes.

2026-10-25 · SUNDAY · 13:11 PDT

  1. 1 Nanbeige/Nanbeige4.3-33B 0 DOWNLOADS · 297 LIKES Nanbeige released Nanbeige4.3-33B, the flagship teacher model of its Nanbeige4.3 series, under Apache 2.0. The text-generation model targets agentic coding and professional cowork tasks, and its card benchmarks it against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at the same scale. It is also meant to anchor distillation into smaller models such as a planned Nanbeige4.3-3B.
  2. 2 orcarouter/DeepSeek-V4.1-Flash-1017-abliterated-GGUF 7,293 DOWNLOADS · 213 LIKES orcarouter published GGUF builds of its abliterated DeepSeek-V4.1-Flash-1017, aimed at running the mixture-of-experts model locally through llama.cpp. The repository went up on October 22 and has drawn 7,293 downloads and 213 likes. Access is gated, so downloading requires a signed-in Hugging Face account.

2026-10-24 · SATURDAY · 15:23 PDT

  1. 1 linoyts/ltx2-19b-distilled-faceswap-v3 78 DOWNLOADS · 140 LIKES This repo fuses a distilled LoRA and a face-swap LoRA into the LTX-2 19B video transformer, shipping the merged weights as four safetensors shards in diffusers format rather than as an adapter. Both of its commits date from March and set the two LoRA scales at 1.0 and 1.1. It is trending on 140 likes against 78 recorded downloads.
  2. 2 HuggingFaceBio/Carbon-X 570 DOWNLOADS · 144 LIKES Hugging Face Bio released Carbon-X, a 14.8B decoder-only model that reads raw DNA inside an ordinary chat and answers in English, predicting whether variants are pathogenic and which regions control gene activity. DNA and text share one token stream, with DNA encoded as 6-mers. It averages 66.7 percent across five DNA benchmarks, 27 points above the Qwen3-14B it starts from.
  3. 3 briaai/Fibo-2.0-dev 1,024 DOWNLOADS · 137 LIKES Bria published Fibo-2.0-dev, the diffusers release of its Fibo 2.0 image model, tagged for text-to-image, editing, inpainting and multi-reference work and shipping a LoRA fine-tuning script. The weights are gated behind an auto-approved form and licensed for non-commercial use only. Comfy-Org repacked the same model into single-file ComfyUI builds two days later.
  4. 4 FastVideo/FastH3-Omni-Comfy 154,490 DOWNLOADS · 130 LIKES FastVideo published ComfyUI-ready builds of its final FastH3 Omni release, which generates video with synchronized audio from a prompt plus ordered image, video and audio references. The repo ships pruned 8-step diffusion weights in BF16 and INT8 ConvRot, alongside Qwen3-VL text encoders and MiniMax-H3 video and audio VAEs. It has drawn 154,490 downloads since October 20.

2026-10-23 · FRIDAY · 01:08 PDT

  1. 1 MiniMaxAI/MiniMax-M3.1-Flash 357 DOWNLOADS · 136 LIKES MiniMax released M3.1-Flash, a native multimodal mixture-of-experts model aimed at coding, agentic workflows, computer use and cybersecurity, with roughly 428B total parameters and 23B active per token. It accepts text, image and video input across a context window of up to 1M tokens. MiniMax reports 73.7 percent on DeepSWE v1.1, 61.7 on OSWorld V2 and 82.2 on CyberGym.
  2. 2 KRAFTON/Raon-1-mini-preview 249 DOWNLOADS · 136 LIKES Game studio KRAFTON published Raon-1 mini, a 30B-A4B mixture-of-experts model it says was trained from scratch, under the Apache 2.0 license. The preview release has built-in reasoning, tool calling for agent workflows, and training across seven languages with an emphasis on Korean. A DFlash2 draft model ships for speculative decoding, alongside MXFP8 and NVFP4 quantizations.

2026-10-22 · THURSDAY · 13:14 PDT

  1. 1 facebook/dinov3-vitl16-pretrain-lvd1689m 738,000 DOWNLOADS · 1,489 LIKES Meta's DINOv3 ViT-L/16 is the 303-million-parameter checkpoint of its self-supervised vision backbone, pretrained on the LVD-1689M image collection and listing the 7B DINOv3 model as its base. It is published for image feature extraction, producing dense representations for downstream vision work rather than text output. Weights are gated and the repo has drawn 738,000 downloads and 1,489 likes.

2026-10-21 · WEDNESDAY · 13:11 PDT

  1. 1 InternScience/Sage-A1.5 58 DOWNLOADS · 140 LIKES Shanghai AI Laboratory released weights for Sage-A1.5, a 35B Mixture-of-Experts agentic model built on A1 for long-horizon scientific work. The card highlights subagent orchestration, verification and open scientific discovery, and tags the model for vision, code and tool use. Apache 2.0, with 58 downloads and 140 likes so far.
  2. 2 RecursiveSI/flint 0 DOWNLOADS · 131 LIKES Recursive published Flint, a mixture-of-experts model for coding agents with 39.8 billion total parameters and about 2.5 billion active per token. The card reports 54.9 percent on Terminal-Bench 2.1, 89.7 percent on SWE-bench Pro v2 and 74.2 percent on SWE-bench Verified. This repo holds the BF16 checkpoint and an SGLang serving kit; it has 131 likes and no downloads yet.
  3. 3 XHToken/Spark-X2.6-4B 85 DOWNLOADS · 132 LIKES Spark-X2.6-4B is XHToken's compact on-device agent model, arriving about a month after Spark-X2.5-4B. It keeps the earlier architecture unchanged and credits improved post-training for its gains in coding, agentic use, long context, multilingual work and reasoning. Apache 2.0, English and Chinese, with 85 downloads and 132 likes.
  4. 4 ornith-ai/Ornith-1.5-122B 742 DOWNLOADS · 125 LIKES Ornith-1.5-122B is a 122B mixture-of-experts model whose self-improvement loop now jointly optimizes task generation, scaffold construction and solution rollouts instead of relying on human-curated tasks and hand-designed harnesses. The team positions it for efficient single-GPU deployment, with a quantized Mobile variant for edge devices. MIT licensed, 742 downloads and 125 likes.

2026-10-20 · TUESDAY · 13:11 PDT

  1. 1 lodestones/Kromoss 5,365 DOWNLOADS · 162 LIKES Kromoss is a full fine-tune of OpenMOSS-Team's MOSS-Image-Base at 1024 pixels, released as a drop-in replacement that keeps the same architecture, Qwen3.5-9B text encoder and FLUX.2 decoder as MOSS-Image-SFT. The repo also ships two 8-step turbo LoRAs distilled from the base model. It has about 5,400 downloads and 162 likes.
  2. 2 inclusionAI/Ling-3.1-flash 2,905 DOWNLOADS · 139 LIKES InclusionAI released Ling-3.1-flash, a hybrid-linear mixture-of-experts reasoning model with 560 billion total parameters, 25 billion active per token and a 1 million token context window. The MIT-licensed upgrade to Ling-3.0-flash targets agents, search, office work and software engineering, reporting results on GDPVal-AA V2.1, FrontierSWE and Terminal-Bench 4.0.
  3. 3 orcarouter/OrcaCyber-Nano-1.0 18,606 DOWNLOADS · 148 LIKES OrcaCyber-Nano-1.0 publishes GGUF builds of a cybersecurity-focused vision-language model meant to run locally under llama.cpp. The card lists reverse engineering, reasoning and function calling, and the repo ships quantizations from a 2 GB Q8_0 up to a 15.7 GB mix plus a separate vision projector. Access is gated, and it has about 18,600 downloads.
  4. 4 nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB 6,820 DOWNLOADS · 135 LIKES NVIDIA released a natively 2-bit build of Nemotron 3.5 Bolt, a vision-language model with 74 billion total parameters and 9 billion active per token, quantized to fit in 24 GB of memory. NVIDIA positions it for everyday agentic work on a local machine, running alongside other applications under the OpenMDW-1.1 license.