xClean Tools

AI Models — Daily Top 5

UPDATED 2026-11-02 01:15 PST

2026-10-05 · MONDAY · 13:12 PDT

  1. 1 autotrust/JEV-27B-VL 1,278,569 DOWNLOADS · 435 LIKES JEV-27B with vision: a LoRA on Qwen3.8-27B that answers typed questions over text and images - yes or no, one of 2 to 256 options, or a rating from 0 to 5 - returning a calibrated probability for every option in a single forward pass. An October 3 update turns those decisions into control, reporting 75 percent of 20 simulated robot arm pick-and-place scenes and 95 percent of 60 multi-step browser tasks completed.
  2. 2 autotrust/GEV-26B-Decide 446,527 DOWNLOADS · 347 LIKES A decision model with adaptive thinking built on Gemma-4-26B-A4B-it, 26 billion parameters with about 4 billion active per token. The card reports a Decision Index 0.2.1 balanced-skill score of 62.48 with adaptive thinking on, against 57.91 for TypeSafe's Jev 1.13 on the same board.
  3. 3 jialinyyzz/humanizer 10,329 DOWNLOADS · 197 LIKES A 12B model that rewrites AI-written drafts - emails, essays, reports and forum posts, in English and Chinese - so they read as if a person wrote them. It is trained to keep every number, unit, date, name and quote and to add nothing, and it runs locally. The author states that no AI detector was used anywhere in training.
  4. 4 deepseek-ai/DeepSeek-V4.1-Flash 869,321 DOWNLOADS · 4,114 LIKES A multimodal Mixture-of-Experts model with a 552B-parameter backbone and support for contexts up to one million tokens, processing images and text natively. Its Causal Encoder-Decoder design splits a 40-layer Transformer into a 20-layer causal encoder followed by a 20-layer decoder, with the decoder's global KV cache projected from the final encoder hidden states.

2026-10-04 · SUNDAY · 13:04 PDT

  1. 1 facebook/sam3 2,162,645 DOWNLOADS · 3,749 LIKES SAM 3 is Meta's Segment Anything checkpoint for mask generation, distributed in safetensors for the transformers library with a video variant in the same repository. The weights are gated and need an approved Hugging Face account to download. It leads the day's pool with about 2.16 million downloads and 3,749 likes.
  2. 2 nvidia/Kumo-Tabular 0 DOWNLOADS · 106 LIKES Kumo Tabular is NVIDIA's pretrained tabular foundation model for classification and regression on structured data. It takes labeled rows as in-context examples and predicts class probabilities for new rows, served through the structured-data-models package with an ensemble of estimators. The weights are released under OpenMDW 1.1.
  3. 3 microsoft/FrogNano-4B-2609 408 DOWNLOADS · 98 LIKES FrogNano is Microsoft's compact repository-level coding agent, derived from Qwen3.5-4B and post-trained only with reinforcement learning on roughly 1,500 synthetic software-engineering tasks. It operates through a five-tool harness called Leaf with an evaluated interaction context of about 131K tokens. Microsoft says the post-training used no solution trajectories or reasoning traces from a stronger model.

2026-10-03 · SATURDAY · 13:09 PDT

  1. 1 Aleph-Alpha/Kolibri-1 0 DOWNLOADS · 209 LIKES Aleph Alpha released Kolibri 1, an Apache 2.0 mixture-of-experts reasoning model with 78B total and 3.46B active parameters, focused on German and English. It has 50 layers of 384 experts, an explicit reasoning mode and tool calling, and ships as FP8 weights with context validated to 1,048,576 tokens. Pre-training ran over 20T bilingual tokens on 768 B200 GPUs.
  2. 2 Infatoshi/GLM-5.3-UNCENSORED-EXL3-3.0bpw 520 DOWNLOADS · 171 LIKES An EXL3 quantization at 3.04 bits per weight packs a weight-edited, uncensored variant of GLM-5.3 - 753B parameters across 256 routed experts - into 273 GiB. Measured against its FP8 source on wikitext-2, KL divergence is 0.089 and perplexity 3.440 against 3.302. On tau2-bench it scores 0.640 on airline and 0.482 on retail, with neither gap against stock GLM-5.3 statistically significant.
  3. 3 lilylilith/QI_2.1_AnyAngle 0 DOWNLOADS · 116 LIKES This LoRA for Qwen-Image-2.1 transfers an arbitrary camera angle onto an existing image without the style drifting. The workflow turns the source into a Gaussian splat or 3D model, re-aims a camera in Blender, and feeds that coarse render back as a reference alongside the original. Training paired real Blender renders with MiniMax H3 orbit videos so illustration and sketch styles hold up too.

2026-10-02 · FRIDAY · 13:13 PDT

  1. 1 pablodawson/MiniMax-H3-360-Orbit-LoRA 3,534 DOWNLOADS · 133 LIKES A LoRA for the first-last-frame variant of MiniMax-H3 that turns a single photo into a frozen-time 360 degree camera orbit. Passing the same image as both keyframes makes the clip close on the frame it started from, so orbits can be chained or stitched into longer shots without a visible seam. Only the camera moves; people, objects and background stay in place.
  2. 2 NaiveAI/Naive-N0.5-Flash 1,365 DOWNLOADS · 130 LIKES Naive-N0.5-Flash is an open-weight 309B mixture-of-experts model with 15.5B active parameters, built for coding and AI research work. It reaches a native one-million-token context with no full-attention layers at all, mixing 39 sliding-window layers with 9 DeepSeek sparse-attention layers over a MiMo-V2.5 base. Weights and inference code are MIT licensed.
  3. 3 zai-org/GLM-5.3 1,413,372 DOWNLOADS · 2,058 LIKES GLM-5.3 shares its base model with GLM-5.2, so every gain comes from post-training. Z.ai reports a 50 percent improvement on its in-house code benchmark and open-source state of the art on Terminal Bench 3.0, where it scores 28.3 against GLM-5.2's 4.6, and on Agents' Last Exam. Cyber capability also emerged as post-training scaled, topping CyberGym for vulnerability discovery.

2026-10-01 · THURSDAY · 13:12 PDT

  1. 1 Cloudflare/clef 18 DOWNLOADS · 172 LIKES Cloudflare released Clef, a 27B multimodal model that takes a state plus a schema of typed questions and returns a probability for every allowed option in a single forward pass. The state can be text, JSON, images or video. The card links a Cloudflare blog announcement and a Decision Index leaderboard.
  2. 2 FermionResearch/Phonon-2 1,325 DOWNLOADS · 135 LIKES An English speech recognition model for Apple Silicon that the card calls the most accurate open model under 900 MB, averaging 5.21 percent word error across the seven English sets of the Open ASR Leaderboard. Quantization-aware training holds each encoder weight at one of five learned levels, keeping the accuracy of a 2.5 GB teacher.
  3. 3 PrunaAI/Pruna-Qwen-Image-2.1 98,442 DOWNLOADS · 133 LIKES A set of LoRA adapters that cuts Qwen-Image-2.1 to five or eight sampling steps for both generation and editing, with no classifier-free guidance and up to 6.3x faster inference. The adapters load on top of the base model, so the pipeline, text encoder and VAE stay unchanged. Training is based on distribution matching distillation.

2026-09-30 · WEDNESDAY · 13:10 PDT

  1. 1 prism-ml/Ternary-Bonsai-2-27B-gguf 3,676,692 DOWNLOADS · 2,297 LIKES Prism ML published Bonsai 2 27B with ternary transformer weights, packaged for llama.cpp on CUDA, Metal and CPU. The card puts the language model at about 5.9 GB against roughly 54 GB at FP16, retains what it calls 98.2 percent of FP16 intelligence as an 84.78 average across 14 benchmarks, and reports about 47 tokens per second on an Apple M5 Max laptop. It has passed 3.6 million downloads.
  2. 2 PSRben/VisionHOPE 137 DOWNLOADS · 221 LIKES The official pretrained weights for VisionHOPE, the paper presenting a visual backbone formulated as a self-modifying learning system. The MIT-licensed repository holds hierarchical VisionHOPE-T, -S and -B checkpoints for ImageNet-1K classification, COCO detection and instance segmentation, and ADE20K semantic segmentation. A day after release it has 221 likes against 137 downloads.

2026-09-29 · TUESDAY · 13:08 PDT

  1. 1 akatz-ai/MiniMax-H3-Character-Swap-LoRA 5,890 DOWNLOADS · 150 LIKES An experimental LoRA from Akatz Labs that swaps a chosen character in a video for one supplied as an image reference, built on MiniMax-H3 and trained for 1,000 updates. The authors report it holds the source scene's background more closely than the base model, while motion timing, facial expressions and hard cuts stay unreliable.
  2. 2 orcarouter/OrcaSAQ-2-Cyber-27B-Uncensored-GGUF 6,170 DOWNLOADS · 150 LIKES A gated, mixed-precision GGUF for llama.cpp, and the cyber-focused entry in orcarouter's OrcaSAQ-2 quantization line. It is built on the maintainer's own uncensored Qwen3.8-27B fine-tune rather than the stock Qwen weights, so it carries that tuning and is not only a change of format.
  3. 3 ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF 19,505 DOWNLOADS · 132 LIKES ISTA-DASLab removed half the experts from Qwen3.8-Flash-Next's 512-expert mixture and quantized what remained with its GSQ and RCO methods, targeting code while keeping the model's multimodal input. The result is 58.4 GB against 354 GB at BF16, an effective 1.89 bits per parameter of the original transformer.