We're sharing/showcasing best of @huggingface models. Follow to stay in loop. Promoting Open-Source models.
Joined December 2025
- Tweets31.8K
- Following32
- Followers71.8K
- Likes624
With 2.3K downloads and growing, Beam-1.0 is proving that sparse MoE plus FP8 is the sweet spot. Top-tier reasoning and code performance without the usual hardware headaches. Try it for your next project.
Mixture-of-experts architecture with FP8 precision keeps it fast and memory-friendly. 501B total parameters but just 23B active, so you get big-model quality with small-model speed. English-focused and ready for real workloads.
Use it for complex reasoning chains, coding assistants, and tool-augmented agents. The conversational and text-generation tags mean it handles chat and long-form content too. Build smarter apps without massive compute.
Meet Beam-1.0-501B-A23B-FP8: a 501B parameter MoE model with only 23B active per token. Built for reasoning, code, and tool use. This is the kind of efficiency that makes big models practical.
With 2,156 downloads and growing, void.1 proves its value. The byte-level tokenization and Qwen3 backbone offer robust performance for text generation, making it a top-tier choice for developers.
Trained with transformers and tensorboard, void.1 uses safetensors for efficient loading. The Qwen3 architecture and 400B token climbmix dataset give it a strong foundation for English text generation.
Use void.1 for creative writing, code generation, and complex text tasks. Its byte-level approach handles diverse inputs, making it great for building chatbots, content tools, or research prototypes.
Meet void.1 by dotlabs: a text-generation model built on Qwen3 with byte-level pretraining on the massive climbmix-400b-shuffle dataset. It's a fresh take on symbolic reasoning for the community.
With 1,039 downloads and 59 likes, Phonon-40m-en is gaining traction for its compact size and streaming TTS. It proves you don't need huge models for quality speech synthesis, especially when running locally on edge devices.
Phonon-40m-en is a 40M parameter transformer for TTS, available in safetensors and GGUF formats. It's designed for efficient inference, with streaming support and on-device deployment in mind, making it lightweight yet capable.
Use Phonon-40m-en to add voice to apps, build accessible reading tools, create real-time narration, or power voice assistants. It's a text-to-audio transformer that outputs speech directly, perfect for streaming and on-device TTS.
Meet Phonon-40m-en: a tiny but mighty text-to-speech model that runs on-device and streams audio in real time. At just 40M parameters, it brings natural voice synthesis to phones and edge devices without the cloud.
1,563 downloads and 59 likes show growing interest. The value? Faster inference without sacrificing quality. If you're running Qwen3.8-Flash-Next, this draft model is your shortcut to lower latency.
Built on Qwen3 architecture with block diffusion and DFlash2. Ships in safetensors for transformers. It's a draft model, so it's lightweight and designed to complement the main Qwen3.8-Flash-Next model.
Use it as a draft model in speculative decoding setups. Works with SGLang for fast inference. Build low-latency chat apps, real-time assistants, or any text generation pipeline where speed matters.
