Hugging Face posted Welcome Beam on October 16, the official changelog for Reflection's 501 billion parameter Apache-2.0 coding model. Unsloth followed later that day with a Dynamic GGUF of those weights, and OpenMOSS Base and SFT of MOSS-Image moved onto the default trending list. A 3.67 GB Q4_K_M GGUF of FLUX.3 Image Klein from Unsloth is on the Hub, and NVIDIA's Qwen3.8-Flash-Next-FP4-Local pack, quantized for DGX Spark and a single RTX PRO 6000, climbed onto that same trending page.
On the Hub
Reflection Beam 1.0
Welcome Beam landed on the Hugging Face blog on October 16. The BF16 card has been on the Hub since October 7: a 501 billion parameter Mixture of Experts with 23 billion active, Apache-2.0, ungated, text in and text out, with a 1 million token context aimed at coding and agentic work. Unsloth's Dynamic GGUF of the same family appeared on October 16; the smallest pack is UD-IQ1_S at 103.06 GB. Reflection's own benches put BF16 Beam at 80.1 on Terminal Bench v2.1, 81.6 on SWE-bench Verified, and 90.6 on GPQA Diamond.12
OpenMOSS MOSS-Image
OpenMOSS's MOSS-Image Base and SFT sit at 26 and 27 on the default trending list this morning. Both cards are Apache-2.0 Diffusers text-to-image packs of 27.229 GB, last written on October 16 after the models landed on October 11. Default output is 1024 by 1024, the bundled VAE is the FLUX.2 small decoder, and a layout-to-image path takes bounding boxes.34
Unsloth FLUX.3 Klein GGUF
Unsloth published a GGUF of FLUX.3 Image Klein on October 16, a local quant of the distilled image model. The Q4_K_M file is 3.667 GB; the tree also carries Q8_0 at 6.093 GB and BF16 at 11.097 GB. A matching Unsloth safetensors pack landed later the same day.56
NVIDIA Qwen3.8-Flash-Next-FP4-Local
NVIDIA's Qwen3.8-Flash-Next-FP4-Local is 30th on the default trending list this morning. The recursive tree is 102.262 GB of NVFP4 tensors, created October 8 and last written October 15, under the NVIDIA Open Model Agreement. NVIDIA validated the pack on a single RTX PRO 6000 Blackwell and on DGX Spark. Upstream vLLM still needs the pinned community containers for this scale-factor format.7
