Skip to content

Support FastFlowLM model downloads from HuggingFace #3152

Description

@korgano

Feature Description

HuggingFace users such as Atomic-Germ are converting models to use NPU faster than the FastFlowLM team. However, Lemonade + FastFlowLM does NOT support downloading and installing Q4NX models from anywhere but FastFlowLM's source.

Use Case / Motivation

  • Increases the number of NPU utilizing models.
  • Increases utility of NPU on NPU equipped APUs.
  • Improves performance of certain workflows depending on model availability.

Platform Relevance

All platforms

Additional Context

Attempting to download any Q4NX model from Atomic-Germ's HuggingFace generates the following:

2026-08-15 08:21:47.118 [Info] (ModelManager) Pulling FLM model: Atomic-Germ/NuExtract3-4B-NPU2:model.q4nx
2026-08-15 08:21:47.119 [Info] (ProcessManager) Starting process: "C:\Users\Pro User\.cache\lemonade\bin\flm\npu\fastflowlm-windows-c3825404ea7c20b3a38c775d761d564254e08925\flm.exe" "pull" "Atomic-Germ/NuExtract3-4B-NPU2:model.q4nx"
2026-08-15 08:21:47.154 [Info] (FLM) [ERROR]  Model not found: Atomic-Germ/NuExtract3-4B-NPU2:model.q4nx; Please check with `flm list` and try again.

Despite failing to download, the models are added to user_models.json:

  "NuExtract3-FFLM": {
    "checkpoint": "Atomic-Germ/NuExtract3-4B-NPU2:model.q4nx",
    "labels": [
      "custom",
      "vision"
    ],
    "recipe": "flm",
    "suggested": true
  },

Metadata

Metadata

Assignees

No one assigned

    Labels

    engine::flmFastFlowLM backend (NPU); multi-modal LLM/ASR/embeddings/rerankingenhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions