We're sharing/showcasing best of @huggingface models. Follow to stay in loop. Promoting Open-Source models.
- Tweets 29,738
- Following 32
- Followers 68,769
- Likes 567
Why it matters: calibrated probabilities mean your agent knows when it's unsure. That's huge for browser and computer use where mistakes cost time. 60 likes and growing. If you're building autonomous agents, this model is a quiet powerhouse.
1
133
Built on Qwen3.5 architecture with safetensors for safe loading. Runs on vLLM for fast inference. The pipeline is zero-shot-classification, meaning no labeled data needed upfront. Designed specifically for decision-making in agentic workflows.
1
34
What can you build? Smarter browser agents that decide what to click, type, or skip. Computer use models that classify UI elements on the fly. Any pipeline needing calibrated probabilities without fine-tuning. Zero-shot means instant deployment.
1
83
Meet openjev: a zero-shot classification model built for AI agents that actually make decisions. It scores and classifies without any training data, perfect for browser and computer use tasks. 1,152 downloads and counting. This is how agents get smarter.
1
3
599
With 8.8k downloads, kev-4b proves its value. Calibration means trustworthy probabilities, and LoRA keeps it lightweight. If you need a solid text classifier that's easy to deploy, this model delivers.
1
179
Under the hood, kev-4b uses Qwen3.5 as its base, fine-tuned with LoRA and PEFT for efficiency. It's calibrated for better confidence scores and typesafe for robust outputs. The model is English-focused and ready for production.
1
35
Use kev-4b for multiple-choice tasks like sentiment analysis, topic tagging, or survey response classification. It's a text classification pipeline, so you can plug it into apps that need fast, accurate decisions. Perfect for building smart filters or recommendation systems.
1
43
Meet kev-4b! A text classification model built on Qwen3.5 with LoRA and PEFT. It's a decision model that's calibrated and typesafe. 8,881 downloads and 62 likes. This one's for anyone who needs reliable multiple-choice predictions.
1
4
696
The big win: abliteration means fewer refusals, so your prompts get encoded as-is. 3,607 downloads show real demand. If you want a text encoder that respects creative freedom in Qwen-Image workflows, this is a top pick.
1
1
273
It's an 8B parameter model, a finetune of Qwen3-VL-8B-Instruct, abliterated to remove refusal behavior. Shipped as safetensors, Apache 2.0 licensed. It keeps the vision-language backbone but focuses on encoding text for image synthesis.
1
63
Use it as the text encoder in Qwen-Image pipelines. It turns your prompts into embeddings for image generation, handling complex scenes and styles. Perfect for building uncensored creative tools, art generators, or any app where prompt freedom matters.
1
85
Meet Qwen-Image-2.1-Text-Encoder-Heretic. It's an abliterated text encoder built on Qwen3-VL-8B. For anyone generating images with Qwen-Image, this unlocks uncensored prompt understanding. 3,607 downloads and 64 likes say people are curious.
1
6
1,114
20k+ downloads and counting. If you want a local LLM that's fast, tiny, and totally unfiltered, this is it. Grab the GGUF, fire up llama.cpp, and enjoy the freedom.
1
289
Built on Qwen3.8 architecture and abliterated for uncensored output. Quantized with PrismML into ternary 2-bit GGUF. That means massive size cuts with surprisingly solid performance.
1
108
This is a text-generation model built for llama.cpp. Run it locally for uncensored chat, roleplay, or creative writing. No filters, no limits. Just you and the model, straight up.
1
122
