OPEN LINK ↗
// 1h agoOPENSOURCE RELEASE
StepFun Step-5 Preview GGUF Lands
Victor Cruz has published day-one GGUF conversions of StepFun’s 600B-class Step-5-Preview on Hugging Face, enabling llama.cpp-oriented local deployment. StepFun advertises 27B active parameters, a 1M-token context window, and vision input. ([StepFun](https://www.stepfun.com/step-5-preview), [Hugging Face](https://huggingface.co/vcruz305/StepFun-5-Preview-GGUF))
// ANALYSIS
The real story is how quickly the community turns frontier-scale weights into runnable artifacts, though “local” still means serious hardware.
- –MoE sparsity reduces per-token compute, but not the enormous memory footprint.
- –The smallest published IQ1_S quant weighs roughly 111 GiB; higher-quality variants exceed 200 GiB.
- –Current llama.cpp support falls back to dense attention, limiting the practical value of the advertised 1M-token context.
- –Multimodal sidecars and model-specific runtime support make this an early adopter release, not plug-and-play infrastructure.
// TAGS
step-5-previewllmopen-weightsmoequantizationinferenceopen-sourcelocal-first
DISCOVERED
1h ago
2026-10-15
PUBLISHED
2h ago
2026-10-15
RELEVANCE
9/ 10
AUTHOR
ViC305