Modelpublic

ajensenwaud/recursant-two-sparks-deepseekv4.1-flash-runtime

sourceHugging Faceotherupdated 10d agoView on Hugging Face

0likes

Model Card

Recursant two-Spark DeepSeek V4.1 Flash — prebuilt runtime

Digest-verified Docker archive of the existing GPU-validated Linux ARM64 image for two NVIDIA DGX Spark/GB10 systems. This repository contains runtime software, not model weights. It is not an official NVIDIA, DeepSeek, vLLM or upstream-plugin release.

The archive was produced with docker image save from dsv41-vision:local-prebuilt-20260916, not by committing or exporting a live serving container. No runtime rebuild was performed for publication.

Install

Use the installer and its immutable release metadata at https://github.com/ajensenwaud/recursant-two-sparks-deepseekv4.1-flash . Download this archive at the exact HF revision recorded in that release, verify its size and SHA-256 against artifact-manifest.json, then docker load --input dsv41-vision-linux-arm64.docker.tar.gz. Docker supports gzip-compressed input. Validate linux/arm64, Config, RootFS diff_ids and the source-manifest label against image-identity.json before use. Image IDs can differ between Docker storage backends; the archive config digest is not a registry manifest digest.

Model weights are separate: https://huggingface.co/ajensenwaud/recursant-two-sparks-deepseekv4.1-flash/tree/d032e578f9ed3724e24239a79d408772620b1d9d . Their MIT license is unchanged.

Corresponding source, modifications and notices

Download runtime-sources-and-notices.tar.gz alongside the binary. It contains the retained plugin and ExLlamaV3 source/build inputs from /opt/src, installed vLLM/vllm-exl3 Python source, runtime patches, and upstream notices from the saved image. The modified installed plugin Python file is under usr/local/lib/python3.12/dist-packages/vllm_exl3/exl3.py; the unoverlaid upstream checkout is under opt/src/vllm-exl3. The installed file plus checkout/native build setup and included headers describes the shipped modified plugin, not the unmodified checkout alone. runtime.Dockerfile documents the integration build steps. Source is available here at no additional charge via the same download mechanism as the binary (AGPL section 6(d)). Keep these directions next to any redistributed copy. Operators of modified AGPL network services must also provide the source offer required by section 13 to their users.

See LICENSES-AND-SOURCE.md for scope. This is not a reproducible-full-build claim: the inherited day-zero vLLM wheel's precise upstream source revision remains unknown. Its installed Apache-2.0 Python code is supplied, but its complete native build source is not represented as recovered. Apache-2.0 does not itself require complete corresponding source for object-code redistribution. Original integration is AGPL-3.0-only; each upstream component retains its own terms. The entire container is not relicensed AGPL.

Modifications by this integration (September 2026): disk-backed Engram support, ARM64 compatibility, vision configuration/sparse-indexer changes, native EXL3 geometry eligibility for the verified two-Spark workload, and deployment controls. This software contains source code provided by NVIDIA Corporation.

NVIDIA components remain subject to the included NVIDIA terms, including NGC-DL-CONTAINER-LICENSE. The archive is a derived application runtime with additional integration functionality, not a standalone redistribution of the unmodified NVIDIA base. No NVIDIA endorsement is implied. See the full retained licenses before redistributing or operating it.