xClean Tools

AI Models — Daily Top 5

UPDATED 2026-11-02 01:15 PST

2026-10-19 · MONDAY · 13:11 PDT

  1. 1 mistralai/Timestral-209M 2,601 DOWNLOADS · 183 LIKES Mistral released Timestral, a family of time series foundation models for zero-shot forecasting; this repository holds the 209M checkpoint, with 5M, 28M and 123M versions alongside. The patch-based transformer handles univariate and multivariate series, past and known-future covariates and 21 quantiles, taking up to 32,768 history steps and forecasting 1,024 in one pass. It reports 0.658 MASE on GIFT-Eval.
  2. 2 nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 1,779 DOWNLOADS · 137 LIKES NVIDIA released Dory 1.0 4B, a looped latent-recurrent language model that repeatedly applies a shared block of layers to spend more compute per token. The normalized-GPT architecture interleaves sliding-window and full-attention layers and supports context up to 256K tokens in English. It ships in BF16 under the OpenMDW 1.1 license for commercial and non-commercial use.

2026-10-18 · SUNDAY · 13:05 PDT

  1. 1 nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 7,599 DOWNLOADS · 153 LIKES NVIDIA published an NVFP4 build of Nemotron 3.5 Super VL, a vision-language model aimed at reasoning, coding, tool use and document and video understanding. It interleaves Mamba-2, mixture-of-experts and attention layers with a C-RADIO vision encoder, carrying 120B total and 12B active parameters with context up to 1M tokens. The weights come from quantization-aware distillation of the BF16 release.
  2. 2 reflection/Beam-1.0-501B-A23B 1,591 DOWNLOADS · 146 LIKES Reflection released Beam 1.0, a sparse mixture-of-experts text model with 501B total and 23B active parameters and a 1M-token context window. The Apache-2.0 release ships instruction-tuned weights in BF16, FP8 and NVFP4, and targets agentic tool use, software engineering and math and science reasoning. Reflection says it matches comparable open models using about 3x fewer reasoning tokens.

2026-10-17 · SATURDAY · 13:11 PDT

  1. 1 deepseek-ai/DeepSeek-V4.1-Flash-1017 0 DOWNLOADS · 862 LIKES DeepSeek published DeepSeek-V4.1-Flash-1017, an MIT-licensed refresh of V4.1-Flash that it calls a minor update but which lifts every benchmark on its own card: Terminal-Bench 4.0 goes from 31.2 to 42.3, Apex Bench from 64.6 to 72.1 and ProgramBench from 20.3 to 29.1. DeepSeek says it also finishes tasks faster and with shorter outputs. Weights ship in FP8 with FP4 experts.

2026-10-16 · FRIDAY · 13:14 PDT

  1. 1 datalab-to/indra-ocr 442 DOWNLOADS · 124 LIKES Datalab released Indra OCR, a 1M-parameter recognizer that it says reads up to 49 pages per second on CPU and turns PDFs and images into markdown, plain text or JSON with boxes. The release ships as about 18 MB of ONNX models with no PyTorch, covering line detection, reading order, layout labels and 14 languages.

2026-10-15 · THURSDAY · 13:15 PDT

  1. 1 black-forest-labs/flux-3-image-klein 3,241 DOWNLOADS · 161 LIKES Black Forest Labs opened the weights of FLUX 3 Image Klein today, a 5.6B-parameter image generation and editing model. It keeps the native 4K output, bounding-box placement and pixel-precise editing of the larger FLUX 3 Image, runs on a single consumer GPU, and ships base, distilled and FP8, NVFP4 and INT8 builds. The FLUX Kommunity License allows free commercial use below $5 million in annual revenue.
  2. 2 meta-models/Muse-Spark-Open-816B-A42B 402 DOWNLOADS · 150 LIKES Meta published the weights of Muse Spark Open, a sparse mixture-of-experts model with 816B total and about 42B active parameters and a 1M-token context. Meta aims it at coding and agentic work, calls it the most capable US open-weight model to date, and says it is the first of its class to run on a desktop, on two Apple M5 Ultra Mac Studios. It ships under the Muse Model License.
  3. 3 autotrust/GLM5.3-Flash-E192-DGX-Spark 7,230 DOWNLOADS · 179 LIKES An unofficial build of GLM-5.3-Flash keeps 192 of the 288 routed experts in each layer, selected by neural architecture search, and quantizes them to NVFP4. That cuts the weights to 132 GB from 306 GB while still activating 18B parameters per token, so a single 180 GB Blackwell GPU runs it with about 32 GiB of KV cache. An NVFP4 draft model for speculative decoding ships with it.
  4. 4 Cloudflare/clef-omni 2,949 DOWNLOADS · 121 LIKES Cloudflare released Clef-Omni, a 30B-A3B mixture-of-experts model post-trained from Qwen3-Omni-30B-A3B-Instruct that turns a state and a schema of typed questions into decisions. It reads the state as text, JSON, images, audio or video and returns a probability for every allowed option in a single forward pass, with no free-form text to parse. A joint schema head scores all options of all questions together.

2026-10-14 · WEDNESDAY · 13:11 PDT

  1. 1 JetBrains/Mellum2.1-12B-A2.5B-Thinking 1,783 DOWNLOADS · 130 LIKES JetBrains released Mellum2.1 Thinking, an Apache 2.0 coding model with 12B total and 2.5B active parameters and a 131k context. The architecture is unchanged from Mellum2: almost all the work went into reinforcement learning, including training inside real repositories with shell and file-editing tools. JetBrains reports LiveCodeBench v6 at 82.0 against 69.4 for its predecessor, and 47.0 on SWE-bench Verified.

2026-10-13 · TUESDAY · 13:16 PDT

  1. 1 AtomicChat/Qwen-Image-2.1-Turbo-Uncensored-GGUF 208,575 DOWNLOADS · 402 LIKES AtomicChat published an abliterated build of Qwen3-VL-8B-Instruct, the text encoder Qwen-Image-2.1 runs on, packaged as GGUF to drop into stable-diffusion.cpp beside the Turbo denoiser weights. The uploader reports the encoder's refusal rate falling from 88.9 to 1.2 percent in English and 38 to 0 percent in Russian on held-out prompts, with MMLU unchanged at 77.35 percent.
  2. 2 dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 78,622 DOWNLOADS · 606 LIKES dealignai's build of DeepSeek-V4.1-Flash removes the model's refusal behaviour at the weight level while leaving the native FP8 checkpoint, routed experts, vision tower and 1M-token context in place, so it loads like the base model with no runtime hooks. The underlying mixture of experts has a 552B-parameter backbone and activates 8B to 16B parameters per token.
  3. 3 vllm-sr/d3 1,156 DOWNLOADS · 135 LIKES The vLLM Semantic Router project released d3, a 26B multimodal decision model that answers choice, yes/no and score questions about text, images or video in a single pass and returns a probability for every answer without generating text. It reports 66.3 on the Jev Decision Index 0.3.2, 10.4 points above its 27B predecessor, and a median 54.4 ms per text request. Apache-2.0.