Community quantizations of two Qwen base models now dominate Hugging Face's trending list
Hugging Face's fastest-growing models by download count this week (2026-09-28 to 2026-09-29, per machine.net's daily tracker) are overwhelmingly not new base models at all. Of the fifteen repos with the largest percentage download gains, at least twelve are third-party GGUF quantizations or fine-tunes of two underlying releases: Alibaba's Qwen3.8-27B and Qwen-Image-2.1. Individual entries range from a 216.6% single-day gain (PrunaAI's Pruna-Qwen-Image-2.1 quantization, 13,262 to 41,994 downloads) down to a 34.7% gain, with community accounts like ukisai, bottlecapai, and Jackrong each shipping their own variant.
The pattern holds on the datasets side too, where the fastest-growing entry, XiaomiMiMo/MiMo-V2.6-RL-oss (+77.2%), is a training dataset released alongside Xiaomi's own MiMo-V2.6-Flash-RL model, itself among the week's raw-download-gain leaders at +37.4%. Meanwhile Ollama's own trending list shows the opposite kind of movement: the platform's fastest-growing repos by pulls are almost entirely small percentage gains (2-10%) spread across a wide, more mixed set of established models, from devstral to phi to moondream, rather than concentrated around a single family.
What the concentration signals
Downloads-based trending metrics on Hugging Face reward whatever the community is actively remixing, not necessarily what a lab just shipped — a distinction that matters for reading the platform as a signal of where real capability is advancing versus where quantization tooling and GGUF conversion pipelines have simply made redistribution easy. Qwen's base models earning this much derivative activity is itself notable: it suggests Alibaba's releases are functioning as a substrate other teams build on, similar to how Meta's Llama family anchored the open-weight ecosystem in prior years. Separately, OpenRouter's usage data for the same period shows DeepSeek's deepseek-v4.1-flash leading model consumption at 14.0% of tracked token share across 578 models — a reminder that download-count trending and actual inference volume are two different measures of relevance, and neither alone tells the full story of what's being deployed.