Instructions to use nsalerni/loudink-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use nsalerni/loudink-v1 with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir loudink-v1 nsalerni/loudink-v1
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
loudink-v1 Flowcast client contract (PR13)
External dependency: Flowcast/GemmaFlow client must support dual packages before canary traffic.
Package selection
| Field | Type | Description |
|---|---|---|
model_package |
enum | v5_lite (default production) | loudink_v1 |
kill_switch |
bool | Force v5_lite regardless of remote config |
Manifest fields the client must read
From artifacts/manifests/loudink_v1.json / published manifest.json:
runner_kind:loudink_v1writer_model_config/ adapter paths /writer_chat_template_familyir_model_config/ adapter paths /ir_chat_template_familypolish_mode(defaultminimal)structural_polish_on_markerslearned_router_pathtranscript_first_ir,intent_fast_path(production true)hot_download_mb_primary/hot_download_mb_hard
Telemetry (optional but recommended)
| Event field | Description |
|---|---|
finish_reason |
Include polish/IR/intent tags |
neural_text_present |
bool — whether raw draft was produced |
polish_mode |
effective mode after markers |
package_version |
loudink-v1 semver |
latency_ms |
e2e |
model_invoked |
false when transcript polish/IR/intent FP skipped models |
Rollback
- Set
model_package=v5_liteorkill_switch=true. - Client must keep last-known-good v5-lite package on disk for offline rollback.
- Training repo does not own client release; this file is the interface contract.
Canary traffic gate
Do not route production users to loudink-v1 until:
gate_loudink_v1 --tier canary --gate-mode enforceis green- Client dual-package flag is shipped
- Package SHA256SUMS verified