axiomofmind/MoSA-Sparse-Network

MoSA Sparse Network

MoSA Sparse Network lets one local AI application coordinate specialized language, retrieval, vision, and OCR models. It keeps routine work on smaller models and loads larger specialists only when a task needs them, making private AI workflows practical on a single workstation with limited VRAM.

This Hub repository is the model-source companion to the MoSA Sparse Network application. It records the upstream repositories and pinned revisions used by the released network. It does not contain model weights.

Model sources

This repository does not redistribute third-party model weights. MoSA downloads selected artifacts directly from their official upstream repositories at pinned revisions. Each model remains subject to its original license and usage terms.

MoSA role	Upstream model	Pinned revision	Status
Fast dispatcher	LiquidAI/LFM2.5-1.2B-Instruct-GGUF	047e06635fbe71469926b35ea414537245218200	Admitted
Default worker	unsloth/Qwen3.5-4B-GGUF	e87f176479d0855a907a41277aca2f8ee7a09523	Admitted
Critic	nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF	ba223d14e45525f7fae81db77ea8cabeb2fc6c25	Admitted
Reasoner	Qwen/Qwen3-8B-FP8	220b46e3b2180893580a4454f21f22d3ebb187d3	Admitted
Large solver, Q8	unsloth/Qwen3.8-27B-GGUF	27af057ecb382ddfea5d12837360a8980560e3ed	Admitted
Large solver, Q4	unsloth/Qwen3.8-27B-GGUF	4ca720788d1e01f1bff70c033e0d0028fd02e502	Admitted
Vision specialist	unsloth/gemma-4-E2B-it-GGUF	90f9618340396838ee7ff5b0ba2da27da62953d3	Admitted
Large vision verifier	google/gemma-4-12B-it-qat-q4_0-gguf	29d097773436b69ff9feafd636ab4cf873786537	Admitted
Fast retrieval	microsoft/harrier-oss-v1-0.6b	f9b9dc8d367d443f2479d27aa5d8d2850c0774ee	Admitted
Retrieval baseline	sentence-transformers/all-MiniLM-L6-v2	c9745ed1d9f207416be6d2e6f8de32d1f16199bf	Baseline
Advanced retrieval	google/embeddinggemma-2	914f7f89142e33e77833254d9c9b90c3cef7303b	Admitted
Decision router	LiquidAI/d1-3B	051bcc464b01b9f92942b364d9586b0ef5912432	Admitted
Fast OCR	PP-OCRv6 medium detector PP-OCRv6 medium recognizer	4236c2b61741a259c091fd879dcc4edc339e916c 024cad6a831de75c2c3c26e711ba8c4a82ccd24b	Admitted
Complex OCR	PaddlePaddle/PaddleOCR-VL-1.6	c5630abae1d940eafe0697512a0325494b02ab42	Admitted

The table mirrors the application's versioned configs/models.yaml registry. Admission describes how a model is integrated into MoSA; it is not a general claim about the quality or safety of the upstream model.

Downloading the model set

From a checkout of the application repository, preview the smallest-to-largest download order without changing the model cache:

uv run sparse-network models pull-all --dry-run

After reviewing the upstream licenses, download and validate each pinned model sequentially:

uv run sparse-network models pull-all --acknowledge-licenses

Normal startup never downloads model weights. Existing valid files are skipped, and each model remains subject to its upstream license and usage terms.