DeepSeek V4.1 Flash · TensorFold TP4Four DGX Sparks, our TensorFold engine, vision and 1M context.
git clone https://github.com/bertholomus/deepseek-v4.1-tensorfold-tp4-4xgb10Recipes
Builds that can be repeated. Scripts, configs, notes.
Open GitHub ↗DeepSeek V4.1 Flash · TensorFold TP2Two DGX Sparks. Weights: Mia-AiLab EXL3 2.9 bpw (MIT).
git clone https://github.com/bertholomus/deepseek-v4.1-tensorfold-tp2-2xgb10DeepSeek V4.1 Flash 1017 · TP2EXL3 2.9 bpw on two DGX Sparks: verified configuration, measured rows, GB10 memory notes.
git clone https://github.com/bertholomus/deepseek-v4.1-flash-1017-tp2-2xgb10GLM-5.3 · TensorFold TP4Full GLM-5.3 at 3.0 bpw on four DGX Sparks: quant formula, launch scripts, measured numbers.
git clone https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10DeepSeek V4.1 Flash · DSpark TP4SGLang on four DGX Sparks: config, patches, systemd lane, measured evidence.
git clone https://github.com/bertholomus/deepseek-v4.1-flash-dspark-tp4-4xgb10GLM-5.3 Flash · NVFP4 TP4Evidence-backed NVFP4 deployment check on four GB10s.
git clone https://github.com/bertholomus/glm-5.3-flash-nvfp4-gb10-tp4DeepSeek V4 Flash 0731 · Graph-8, four nodesVerified four-GB10 TP4 DSpark deployment and benchmarks.
git clone https://github.com/bertholomus/deepseek-v4-flash-0731-dspark-graph8-4xgb10DeepSeek V4 Flash 0731 · Graph-8, two nodesVerified two-GB10 DSpark deployment and benchmarks.
git clone https://github.com/bertholomus/deepseek-v4-flash-0731-dspark-graph8Clef · low-bit quantsOriginal quants of Cloudflare Clef (27B) for one 24 GB GPU, with a calibration-first evaluation harness.
git clone https://github.com/bertholomus/clef-quant