xClean Tools

AI Models — Daily Top 5

UPDATED 2026-11-02 01:15 PST

2026-10-12 · MONDAY · 13:14 PDT

  1. 1 autotrust/JEV-35B 6,262 DOWNLOADS · 178 LIKES AutoTrust released JEV-35B, a System 1 decision model fine-tuned from Qwen3.5-35B-A3B that scores up to 256 typed options in a single forward pass, with about 3B of its 35B parameters active per token. The lab reports 59.73 on the public Decision Index 0.3 suite against 53.64 for its JEV-27B-VL, and a median request latency of 204ms on one B200 versus 271ms.

2026-10-11 · SUNDAY · 13:09 PDT

  1. 1 LiquidAI/d1-omni-600M 10,489 DOWNLOADS · 105 LIKES Liquid AI published d1-omni-600M, a 587M-parameter decision model built on its LFM2.5-Encoder-350M. Given a state as text or JSON, optionally with images or up to 30 seconds of speech, it returns typed answers to named questions with zero output tokens, read straight from the model's distribution over the options rather than generated. It covers 15 languages and is sized for the edge.

2026-10-10 · SATURDAY · 13:13 PDT

  1. 1 speridlabs/iris-3b 233 DOWNLOADS · 104 LIKES Iris-3B is a 3-billion-parameter text-to-image model that works directly in pixel space, with no VAE or latent decoder, and outputs images of about one megapixel. Sperid Labs frames the same pixel-space prior as a general vision learner: fine-tuned with no architectural change, it also ships monocular depth estimation, image restoration and upscaling. The Apache-2.0 release accompanies arXiv paper 2610.09450.
  2. 2 perplexity-ai/pplx-decider-v1.1-27b 1,807 DOWNLOADS · 100 LIKES Perplexity updated its 27B decision model, which answers a declared set of options through a calibrated decision head in a single pass instead of generating text. The card reports the Decision Index rising from 56.4 to 61.56 on the same Qwen3.8-27B backbone, credited mostly to lifting the causal mask and training on more data. The checkpoint carries a separate readout rather than a full-vocabulary head.

2026-10-09 · FRIDAY · 01:13 PDT

  1. 1 Qwen/Qwen3.8-Flash-Next 1,640,938 DOWNLOADS · 6,048 LIKES Qwen published Qwen3.8-Flash-Next, the post-trained weights for the model in Hugging Face Transformers format. The repository holds a mixture-of-experts, image-text-to-text model the card describes as compatible with Transformers, vLLM, SGLang and TokenSpeed. Third-party GGUF quantizations are already circulating, one of them with more than 3.4 million downloads.

2026-10-08 · THURSDAY · 13:15 PDT

  1. 1 LiquidAI/d1-3B 5,370 DOWNLOADS · 182 LIKES Liquid AI released d1-3B, a 3B decision model built on LFM2.5-VL-3B that takes a state as text, JSON or images plus a set of questions and returns calibrated typed answers in one forward pass with zero output tokens. The card reports 48.57 on Decision Index 0.2.1, the best under 10B and ahead of Decider 35B-A3B at 47.11, and 8 ms per decision on an RTX 4090.
  2. 2 FrancisRing/Prism 0 DOWNLOADS · 124 LIKES Prism is an MIT-licensed video diffusion transformer from Fudan University, Tencent Hunyuan and Zhejiang University that generates video and audio jointly at native 2K. It uses dynamic sparse attention to make training at that resolution tractable. The repo ships model weights and image-to-video sample cases, alongside a paper and project page.

2026-10-07 · WEDNESDAY · 13:14 PDT

  1. 1 canberkkkkkk/ema-lightning 2,724 DOWNLOADS · 254 LIKES EMA Lightning is a Turkish text-to-speech system in two small parts: a 5.6M-parameter DiT acoustic model and a 3M-parameter vocoder, about 34 MB in total under Apache 2.0. Its model card reports 0.92 percent word error rate on Freya-TR-Eval, the lowest of the systems measured, first audio in 3.86 ms and 440 times faster than real time on an RTX 4090.

2026-10-06 · TUESDAY · 13:12 PDT

  1. 1 google/embeddinggemma-2 364 DOWNLOADS · 264 LIKES Google DeepMind released EmbeddingGemma 2, an open embedding model that maps text, code, images, video and audio into one shared 768-dimensional space. It totals 740M parameters: a 270M text model plus modular 170M vision and 300M audio encoders. The Apache 2.0 release covers over 100 languages and is built to run on phones and laptops for on-device search, RAG and classification.
  2. 2 Cactus-Compute/whistle 1,887 DOWNLOADS · 120 LIKES Cactus Compute published Whistle, a speech-to-text model that ships as a single 16.9 MB file and runs on device CPUs with no GPU and no dependencies. It transcribes up to 30 seconds of 16 kHz mono audio in seven European languages, returns per-word timestamps and probabilities, and can emit raw speech embeddings. It reuses the Needle engine, so devices already running Needle need no second runtime.