← Model Cards

GB10SOR LANES · 2026-09-21

Eights

8 Sparks · 4 models · 4 qualified recipes

NixOS

Choose a recipe, stage its weights, and run the command from the matching GB10SOR checkout. Multi-Spark lanes also need the private local cluster configuration described in the repository. Setup guide →

Use the repository quickstart for DGX OS or NixOS. Multi-Spark lanes also require the matching private local cluster configuration.

incoai

GLM-5.3-NVFP4

incoai/GLM-5.3-NVFP4

NixOSOne-command Nix ready
Recipe options
Engine
SGLang 0.5.19
Served context
16,384 tokens
Request limit
1
Parallelism
TP8 / PP1
Checkpoint weights
432.9 GiB
Recipe YAML
Recipe & author credits

Use the exact checkpoint revision and deployment profile. Combining recipes is a planning choice; it does not establish combined compatibility.

Contributors: incoai for the checkpoint; Z.ai and SGLang upstream; GB10SOR local integration

Checkpoint revision: 54e52520606f96b3d9fc84088ad22882a61648ac

Recipe SHA-256: 06b2cabe3fa38ae36ba0d2505be177158c80f6a2eb55619566665fe3043c9633

Exact model card ↗
lukealonso

Kimi-K3-QSRT-K2

lukealonso/Kimi-K3-QSRT-K2

NixOSOne-command Nix ready
Recipe options
Engine
vLLM 0.11.2.dev280+infernal.de04f08.cu133.torch213
Served context
8,192 tokens
Request limit
1
Parallelism
TP8 / PP1
Checkpoint weights
705.5 GiB
Recipe YAML
Recipe & author credits

Use the exact checkpoint revision and deployment profile. Combining recipes is a planning choice; it does not establish combined compatibility.

Contributors: Luke Alonso; QSRT and vLLM upstream; GB10SOR local integration

Checkpoint revision: c0afd8f8f260d1225f3d0ef63aac3bc2682fc513

Recipe SHA-256: ca8d75a943b498b4febf69638f7897566f69ad2e0e04bd39aa6225027ab45034

Exact model/runtime guidance ↗
nvidia

DeepSeek-V4.1-Flash-NVFP4

nvidia/DeepSeek-V4.1-Flash-NVFP4

NixOSOne-command Nix ready
Recipe options
Engine
SGLang 0.0.0.dev1+gda64c5cbb
Served context
16,384 tokens
Request limit
1
Parallelism
TP8 / PP1
Checkpoint weights
491.1 GiB
Recipe YAML
Recipe & author credits

Use the exact checkpoint revision and deployment profile. Combining recipes is a planning choice; it does not establish combined compatibility.

Contributors: DeepSeek model team; NVIDIA ModelOpt/checkpoint team; SGLang and FlashInfer upstream; GB10SOR SM121 and Engram integration

Checkpoint revision: 3431dde3247c13b5957f682b1e3c6fcae2566079

Recipe SHA-256: 222942fab2bc886e5665abf581ebec2bc2b4ce0e97aede8a4cd31cb362b1ac6f

Exact NVIDIA checkpoint and GB300 recipe ↗
nvidia

NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4

nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4

NixOSOne-command Nix ready
Recipe options
Engine
vLLM 0.22.0
Served context
16,384 tokens
Request limit
1
Parallelism
TP4 / PP2
Checkpoint weights
328.1 GiB
Recipe YAML
Recipe & author credits

Use the exact checkpoint revision and deployment profile. Combining recipes is a planning choice; it does not establish combined compatibility.

Contributors: NVIDIA NeMo/Nemotron team; eugr for the community one-command wrapper

Checkpoint revision: 02462641f13d3af838b904f48195b9bb8a1e4ebc

Recipe SHA-256: c3f59c10bc39335d0ef7b562321335f21315c2d3cc9c9d13bd43be406f04cdb9

Exact NVIDIA 4× Spark guide ↗