The Hub is quiet this morning. Four items set the day. Unsloth quantized Qwen's eight step Image 2.1 Turbo denoiser to GGUF. Horizon posted an INT4 MoE runtime measured on a gaming RTX 5070. A Hub write-up runs SmolLM2 on an Xbox Series S in Dev Mode. JetBrains' official Mellum2.1 Thinking GGUF has been on the Hub since Wednesday.
On the Hub
Unsloth Qwen-Image-2.1-Turbo GGUF
Unsloth published a Dynamic 2.0 GGUF of Qwen-Image-2.1-Turbo on the Hub on October 10. It is the official eight step denoiser, packed for Unsloth Desktop, ComfyUI-GGUF, and stable-diffusion.cpp under a qwen-research license. The recommended file is Q4_K_M at 4.20 GB; the pack is the denoiser only, so you still need Unsloth's VAE and a Qwen3-VL-8B encoder beside it. Unsloth's own LPIPS table against a bf16 Turbo reference, run at 1024 by 1024 with eight steps and CFG 1, puts Q4_K_M at 0.153 and Q8_0 at 0.036.12
Horizon RunMap
Lucas Ribeiro published a Hub blog on October 11 describing Horizon RunMap, an independent MoE runtime that keeps INT4 experts compact on the GPU instead of expanding every expert to FP16. On a shared Windows RTX 5070, the author's own numbers put OLMoE-1B-7B-0924-Instruct at 61.26 tokens per second in 5.99 GiB when all 64 experts stay resident and the Grouped T3 kernels are on. That is a measured local MoE story, not a new lab weight.34
SmolLM2-360M on Xbox Series S
Gianluca Mazza published a Hub write-up on October 11 of running HuggingFaceTB/SmolLM2-360M-Instruct on an Xbox Series S. It is INT4 on the Zen 2 CPU through ONNX Runtime GenAI, sideloaded as a UWP package in Dev Mode. The weights have been on the Hub since 2024, and this is not a store app. What matters is the constraint list: AppContainer path walks, disk versus RAM, and DirectML falling back without telling you.56
JetBrains Mellum2.1 Thinking GGUF
JetBrains put an official Apache-2.0 GGUF of Mellum2.1-12B-A2.5B-Thinking on the Hub on October 7. It is a 12B MoE coder with think-blocks, 64 experts and 8 active, and 131k context. The pack includes MXFP4_MOE plus Q4_K_M, Q6_K, Q8_0, and BF16. This is JetBrains' own llama.cpp pack of the thinking coder, not a community conversion.7
