Hugging Face downloads look stalled, not crashing, when mapped to primary models
natolambert · x · 2026-10-14
After matching 2k curated local checkpoints back to their original "primary" model weights, natolambert finds monthly Hugging Face downloads show more of a stall than a clear decline. The chart shows monthly downloads peaking above 700M in mid-2026 before falling back, with China (red) accounting for a large share, the US (blue) second, and the latest month partially projected. A data-driven calibration to claims that open-weight model usage is collapsing — normalized to primary models, it looks more like a plateau.
Related event: HuggingFace Downloads Plateau as Local Model Deployments Surge(4 posts)→
More from Infra
- Agent token spend hits 1-2x payroll as coding agents rewrite engineering norms — nlarusstone · 2026-10-14
- OCP summit packed as engineers queue for 800Vdc and CPO racks — BenBajarin · 2026-10-14
- Run open-weight LLMs and build agents locally with Docker Model Runner: free, dual llama.cpp/vLLM backends — juntao · 2026-10-14
- First PyTorch-CUDA build for Windows on Arm lands as NVIDIA RTX Spark laptops ship Friday — lmoroney · 2026-10-14
- ShipStation built an AI IT help desk on Bedrock AgentCore from years of Slack history — lateinteraction · 2026-10-14
- PhD student weighs used 3090s, modded 3080 20GB, or new 32GB R9700 for local LLM homelab — MrSkullking · 2026-10-14