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.

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.

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.

Meet Qwen3.8-Flash-Next-DFlash2: a draft model built for speculative decoding. It pairs with Qwen3.8-Flash-Next to make text generation faster by predicting tokens ahead of time. Speed boost incoming.

Meet Ming-3.0-tts-0.5B: a compact text-to-speech model that clones voices in multiple languages. With only 0.5B parameters, it delivers high-quality speech generation that's easy to deploy. Perfect for developers exploring TTS!

Meet Qwen3.8-Flash-Next-Gyro-GGUF: an experimental image-text-to-text model that's already pulling 173K+ downloads. It's a GGUF quantized version of Qwen3.8-Flash-Next, built for fast, local multimodal inference. Curious? Let's dive in. 🧵

Meet Nemotron-3.5-Parse: a new vision language model that turns images into structured text. Built for OCR and parsing tasks, it's here to make document understanding smarter and easier.

Meet Spark-X2.6-4B-GGUF: a 4B parameter text-gen model in GGUF format. Built for efficient local inference, it's perfect for running on consumer hardware. 5k+ downloads and 79 likes show the community is taking notice!

Meet flint-GGUF: a quantized text-generation model from unsloth, built on RecursiveSI/flint. With 7.4k downloads and 141 likes, it's a compact, efficient option for local inference. Ready to generate?

What if an LLM could retain and build on past experiences without relying entirely on external memory systems? That's the direction MemT 1.0 is exploring. Built on Gated DeltaNet-2, it uses token-level memory updates to retain and refine information as it processes new inputs. Its Mixture-of-Storage mechanism is designed to scale native memory capacity, enabling the model to carry useful experience forward and apply it to future tasks. At 128K tokens, MemT-1.0-Nano reports 1.86× faster end-to-end inference and 66× lower persistent storage compared to a full-context KV-cache baseline. An interesting direction for AI systems that need to retain and build on experience across interactions. Explore the models here: huggingface.co/collections/M…