全站搜索

AI 模型 84759

• tiny-SmolVLMForConditionalGeneration↗

  trl-internal-testing

• flux-3-image-klein-base↗

  unsloth

• Qwen3.8-Flash-Next-W4A16↗

  VnimanieAI

• MN-Helcyon-GPT-4o-12b-v4.5-GGUF↗

  XeyonAI

• Qwen3.8-27B-ARA-abliterated-NVFP4-MTP↗

  aday777

  This is a multimodal, refusal-ablated (commonly described as "uncensored") W4A4 NVFP4 derivative of the reproducible `trohrbaugh/Qwen3.8-27B-heretic-ara` BF16 checkpoint. It is intended for native Blackwell NVFP4 inference. This artifact passed its text-capability, image-vision, benign refusal-surface, native MTP-acceptance, integrity, and clean-load gates on vLLM 0.23. See `BUILD_MANIFEST.json`, `VALIDATION_REPORT.json`, and `SHA256SUMS` for exact provenance and results. Video tensors/processors are preserved, but video input was not part of the live runtime gate. The live gate used `Qwen3_5ForConditionalGeneration`, native three-token MTP, the FlashInfer CUTLASS NVFP4 kernel, an 8,192-token context, and an RTX PRO 6000 Blackwell GPU. Loaded model memory was approximately 19.53 GiB.

• DeepScaleR-1.5B-Preview↗

  agentica-org

  DeepScaleR-1.5B-Preview is a language model fine-tuned from DeepSeek-R1-Distilled-Qwen-1.5B using distributed reinforcement learning (RL) to scale up to long context lengths. The model achieves 43.1% Pass@1 accuracy on AIME 2024, representing a 15% improvement over the base model (28.8%) and surpassing OpenAI's O1-Preview performance with just 1.5B parameters. Our training dataset consists of approximately 40,000 unique problem-answer pairs compiled from: - AIME problems (1984-2023) - AMC problems (prior to 2023) - Omni-MATH dataset - Still dataset We employ Deepseek's Group Relative Policy Optimization (GRPO), a simplified RL algorithm that extends PPO by: - Normalizing advantage function over all samples generated from the same prompt. - Applying KL divergence regularization on top of PPO's surrogate loss to prevent significant policy drift.

• qwen3.5-plus-2026-02-15↗

  阿里巴巴 · Qwen

  阿里巴巴于2026年2月16日发布了Qwen3.5-Plus模型。该模型的总参数量为3970亿（397B），其中激活参数为170亿 [2-3]。在多项基准测试中，其表现与Gemini-3-pro、GPT-5.2等模型相近。该模型采用了四项核心架构创新。其训练成本与算力部署成本相较于Qwen3-Max有所降低，推理吞吐量在256K长文本下有所提升

• Crow_Qwen9B_OpusUCH↗

  Babsie

• BlackSheep-Mistral-RP-7B-GGUF↗

  bartowski

• Rombo-Org_Rombo-LLM-V3.0-Qwen-32b-GGUF↗

  bartowski