サイト内検索
AI モデル 85021
- Qwen3.6-35B-A3B-PrismaQuant-4.75bit-vllm↗
rdtand
Mixed-precision quantization of `Qwen/Qwen3.6-35B-A3B` produced by **PrismaQuant** — a per-Linear sensitivity-driven allocator that chooses each Linear module's format individually under a total-bit budget. **Why "every layer refracts into a different format":** a naive uniform NVFP4 either leaves disk on the table (keeping everything BF16 "to be safe") or loses quality (quantizing sensitive layers to 4-bit). PrismaQuant measures the actual Fisher-weighted MSE for every (Linear, format) pair and runs a multi-choice knapsack under a total-bit budget, so every bit lives where it buys the most likelihood. This checkpoint uses **three precisions**, selected per-Linear by the allocator from measured sensitivity — not chosen uniformly:
- abdulmannan-01_-_qwen-2.5-0.5b-finetuned-for-sql-generation-combined-dataset-gguf↗
RichardErkhov
- Changgil_-_K2S3-Mistral-7b-v1.43-gguf↗
RichardErkhov
- SongTonyLi_-_OpenELM-1_1B-SFT-D_chosen-HuggingFaceH4-ultrafeedback_binarized-Xlarge-gguf↗
RichardErkhov
- FrogNano-4B-2609-heretic-GGUF↗
richardyoung
- ppo-Huggy↗
Ryukijano
- Wan2.2_I2V_LightX2V_Slow-Mo_Fix↗
rzgar
Wan_2_2_I2V_lightx2v_Distill_260412_rank_128_fp16 originally extracted from lightx2v/Wan2.2-Distill-Models wan2.2_i2v_A14b_lightx2v_4step_720p_260412.safetensors The default ComfyUI Wan2.2 I2V workflow ships with LightX2V LoRAs that tend to turn scenes into **slow motion**: These **slow-mo fix** LoRAs target that problem. More **natural motion** without breaking the **4-step LightX2V** usage pattern and quality. In practice they also feel among the **fastest-looking** LightX2V-style pairs released for Wan2.2 I2V (less drag, less “slow-mo” pacing).
- blip2-itm-vit-g-coco↗
Salesforce
- llama3-8B-slerp-biomed-chat-chinese↗
shanchen
llama3-8B-slerp-biomed-chat-chinese is a merge of the following models using LazyMergekit: * shanchen/llama3-8B-slerp-med-chinese * shenzhi-wang/Llama3-8B-Chinese-Chat
- Assistant_Pepe_8B↗
SicariusSicariiStuff
