전체 검색

AI 모델 2119

• japanese-hubert-base↗

  yky-h

  This is a mirror of japanese-hubert-base, originally released by rinna Co., Ltd. The original model is licensed under the Apache License 2.0. This mirror follows the same license terms. All copyrights remain with the original authors. The model architecture is the same as the original HuBERT Base model, which contains 12 transformer layers with 12 attention heads. The model was trained using code from the official repository, and the detailed training configuration can be found in the same repository and the original paper. The model was trained on approximately 19,000 hours of following Japanese speech corpus ReazonSpeech v1. - ReazonSpeech

• Smaug-Llama-3-70B-Instruct-32K↗

  abacusai

  This is a 32K version of Smaug-Llama-3-70B-Instruct. It uses PoSE (https://arxiv.org/abs/2309.10400) and LoRA (https://arxiv.org/abs/2106.09685) adapter transfer. More details are coming soon. Score vs selected others (sourced from: (https://lmsys.org/blog/2024-04-19-arena-hard/#full-leaderboard-with-gpt-4-turbo-as-judge)). GPT-4o and Gemini-1.5-pro-latest were missing from the original blob post, and we produced those numbers from a local run using the same methodology. Note that we believe the number of tokens/verbosity of the model strongly influences the GPT-4 judge in this case, and at least partially explains the improvement in Arena-Hard score for the 32K model.

• FastVideo-FastH3-8-Step-V2-NVFP4-Comfy↗

  FastVideo

  Original model repository: https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2 FastH3 V2 with its transformer quantized to NVFP4, for NVIDIA Blackwell GPUs: RTX 50 series, RTX PRO 6000 and DGX Spark. It is built from `fastvideo_fasth3_8step_v2_pruned_bf16.safetensors` in FastVideo/FastVideo-FastH3-Comfy, which also has the BF16 and INT8 builds for other GPUs. Use the **FastVideo FastH3: Text to Video** template (ComfyUI 0.36.0 or later) and select `fastvideo_fasth3_8step_v2_pruned_nvfp4.safetensors` (13.6 GB) in its `UNETLoader`. Keep the template's sampler settings: 8 steps, `res_multistep`, `simple`, and the `MiniMaxH3SigmaShift` node at 10 (video) and 3 (audio).

• SDAR-8B-Chat↗

  JetLM

  **SDAR** (**S**ynergy of **D**iffusion and **A**uto**R**egression) model is a new large language model that integrates autoregressive (AR) and discrete diffusion modeling strategies. It combines the efficient training paradigm of AR models with the highly parallel inference capability of diffusion models, while delivering performance fully on par with SOTA open-source AR models. At the same time, SDAR sets a new benchmark as the most powerful diffusion language model to date. We highlight three major conclusions from our study: > [!IMPORTANT] > Take-home message > > - **Balanced Efficiency:** SDAR unifies the **efficient training** of AR models with the **parallel inference** of diffusion, achieving both fast training and inference. > - **Fair Comparisons:** In rigorously controlled experiments, SDAR achieves **on-par general task performance** with strong AR baselines, ensuring credibility and reproducibility. > - **Superior Learning Efficiency:** On complex scientific reasoning tasks (e.g., GPQA, ChemBench, Physics), SDAR shows **clear gains over AR models** of the same scale, approaching or even exceeding leading closed-source systems. For **SDAR** models, inference hyperparameters are set to: `block_length = 4`, `denoising_steps = 4`, greedy decoding.

• sam2.1-hiera-base-plus-ONNX↗

  onnx-community

• nfnet_l0.ra2_in1k↗

  timm

  Normalization Free Networks are (pre-activation) ResNet-like models without any normalization layers. Instead of Batch Normalization or alternatives, they use Scaled Weight Standardization and specifically placed scalar gains in residual path and at non-linearities based on signal propagation analysis. Lightweight NFNets are `timm` specific variants that reduce the SE and bottleneck ratio from 0.5 -> 0.25 (reducing widths) and use a smaller group size while maintaining the same depth. SiLU activations used instead of GELU.

• MOSS-Image_4-step_adapter_lora↗

  darask0

  **English summary.** A 4-step, CFG-free adapter for OpenMOSS-Team/MOSS-Image-SFT (50 steps × 2 forwards with CFG 3). It continues lodestones/MOSS-Image-SFT-8-steps-dmd (data-space DMD, arXiv 2610.12102) restricted to 4 iterations, trained with DMAD (arXiv 2610.02188): a LoRA critic on the frozen teacher separates teacher samples from student samples, and the student minimises the negative critic logit. DiT time per 1024² image on one H100: 16.2 s (teacher) → 0.81 s. Compared with the lodestones checkpoint run for 4 steps, images have more detail and livelier composition, with some regressions (see Results). Same file format and sampler as the lodestones checkpoint; `inference.py` runs it with the release pipeline. Training code is in `code/`. MOSS-Image-SFT(50 ステップ・CFG 3 で 1 枚 100 回の DiT 計算)を、**4 回の計算・CFG なし**で画像を出せるようにする高速化アダプタです。 lodestones の 8 ステップ版(data-space DMD、update 2600)を出発点に、4 ステップに絞って DMAD で追加学習しました。 カスタムノード `comfyui/rapid-moss/` を同梱しています。ComfyUI 本体も MOSS-Image に対応していますが、専用サンプラーと time path の置き換えはなく、標準の `LoraLoaderModelOnly` は LoRA を bf16 の重みに足し込みます(当たる差分の相対誤差 81%)。標準ノードだけでも絵は出ますが、学習した生成器とは別物になります(下の比較)。このノードを使ってください。

• sesame-csm-1b-GGUF↗

  ggml-org

• Qwen3-4B-SafeRL↗

  阿里巴巴 · Qwen

  **Qwen3-4B-SafeRL** is a safety-aligned version of the Qwen3-4B model. It has been trained using Reinforcement Learning (RL) with a reward signal from **Qwen3Guard-Gen** to enhance its robustness against harmful or adversarial prompts. This process aims to ensure strong safety guarantees without leading to overly simplistic or evasive refusal behaviors, thereby maintaining a positive user experience. To prevent the model from defaulting to refusal across all prompts in an attempt to remain safe, we introduce a hybrid reward function that jointly optimizes three key objectives: Qwen3-4B-SafeRL is used in the same way as Qwen3-4B, preserving the ability of hybrid thinking modes. The code of Qwen3 has been in the latest Hugging Face `transformers` and we advise you to use the latest version of `transformers`.

• GLM-4-32B-0414-GGUF↗

  unsloth

  The GLM family welcomes new members, the **GLM-4-32B-0414** series models, featuring 32 billion parameters. Its performance is comparable to OpenAI’s GPT series and DeepSeek’s V3/R1 series. It also supports very user-friendly local deployment features. GLM-4-32B-Base-0414 was pre-trained on 15T of high-quality data, including substantial reasoning-type synthetic data. This lays the foundation for subsequent reinforcement learning extensions. In the post-training stage, we employed human preference alignment for dialogue scenarios. Additionally, using techniques like rejection sampling and reinforcement learning, we enhanced the model’s performance in instruction following, engineering code, and function calling, thus strengthening the atomic capabilities required for agent tasks. GLM-4-32B-0414 achieves good results in engineering code, Artifact generation, function calling, search-based Q&A, and report generation. In particular, on several benchmarks, such as code generation or specific Q&A tasks, GLM-4-32B-Base-0414 achieves comparable performance with those larger models like GPT-4o and DeepSeek-V3-0324 (671B). **GLM-Z1-32B-0414** is a reasoning model with deep thinking capabilities. This was developed based on GLM-4-32B-0414 through cold start, extended reinforcement learning, and further training on tasks including mathematics, code, and logic. Compared to the base model, GLM-Z1-32B-0414 significantly improves mathematical abilities and the capability to solve complex tasks. During training, we also introduced general reinforcement learning based on pairwise ranking feedback, which enhances the model's general capabilities. **GLM-Z1-Rumination-32B-0414** is a deep reasoning model with rumination capabilities (against OpenAI's Deep Research). Unlike typical deep thinking models, the rumination model is capable of deeper and longer thinking to solve more open-ended and complex problems (e.g., writing a comparative analysis of AI development in two cities and their future development plans). Z1-Rumination is trained through scaling end-to-end reinforcement learning with responses graded by the ground truth answers or rubrics and can make use of search tools du