Fine-tuned 350M model for road-freight voice workflows in mobile apps. It is built for app tasks like navigation, load search, quoting, form repair, instruction following, and skill selection.

Intended use

Use this model inside a road-freight app, with the app supplying its tool definitions, current screen state, and confirmation rules. It is not a standalone assistant, route planner, pricing authority, or live operations record.

Give the model live app state: available vehicles, load cards, load details, quote state, current location, and whether the driver has confirmed an action. The app still validates every tool call and keeps final submission behind its own confirmation checks.

Quick start

For the BDAIAssistant SDK, use the .small model family. The SDK defaults that family to Q4_K_M.gguf, the recommended mobile file.

swift

let spec = Lfm25GgufFileSpecification(family: .small, variant: .q4km)
let path = try await Lfm25LlamaCppLLMDownloader().downloadGgufFile(spec)

To download the GGUF directly:

hf download bitsydarel/road-freight-voice-assistant-350m Q4_K_M.gguf

Files

fused/: Hugging Face-format fused model.
model.gguf: F16 GGUF exported from the fused model.
Q8_0.gguf, Q6_K.gguf, Q5_K_M.gguf, Q5_K_S.gguf, Q4_K_M.gguf, Q4_K_S.gguf: available GGUF quantizations.
quantization.manifest.json: SHA-256 hashes and tensor-contract validation results for the GGUF files.

For mobile deployments, start with Q4_K_M.gguf if the package size works for your app. Use Q4_K_S.gguf when size matters more. Keep Q8_0.gguf for comparison checks, and use model.gguf if you need to re-export or quantize again.

This release does not publish BF16.gguf or Q4_0.gguf. The BDAIAssistant SDK treats those variants as unavailable for the .small family and fails before trying to download them.

Validation

The GGUF files listed above were regenerated from the fused model and passed the runtime tensor contract checks, including rank-2 shortconv tensors and token embedding dimensions.

License and base model

This release is a fine-tune of LiquidAI/LFM2.5-350M and follows the upstream lfm1.0 license linked in the repository metadata.