You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Step-5-Preview EXL3 (SAGE)

Continuity + release repo for the SAGE-assisted EXL3 quantization of Step-5-Preview-BF16 — MoE, 92 layers, 352 experts top-8, GQA 64/4, head_dim 192, 262144-token window — targeting 2x DGX Spark (GB10 / sm_121, 128 GB unified each).

Status: work in progress — no weights published yet. This repo tracks the work files (census, plans, calibration manifests, recipes, reports) as they are produced, before encode, so the pipeline can resume from any point.

Access

Public repo with gated access (manual review). Request access and it will be reviewed. Nothing here is a quality claim; see the evidence policy below.

Layout

work/STATUS.json — machine-readable progress
work/census/ — source inventory (arch config, shard sizes, headers)
work/plans/ — stream plans, size targets, budgets
work/logs/ — small logs and manifests only
recipes/ — compiled SAGE recipes
reports/ — measurement and validation reports

Evidence policy

Implemented code, CPU tests, GPU qualification, predicted recipe quality and measured final-model quality are different evidence levels. Never present one as another.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support