loudink-v1 v0.4.0

On-device dual-stack for LoudInk / Flowcast on Apple Silicon:

1. Writer — LFM2.5-1.2B LoRA (dictation polish / messages / email / prompts)
2. Planner — FunctionGemma compact-IR LoRA (computer-use plans)

Predecessor: nsalerni/flowcast-v5-lite

v1.1 track: record-and-replay skills, richer real-world benches. This 0.4.0 freeze is the loudink-v1 GA snapshot.

Daily-Mac product bar (v0.4)

Slice	Shipped	Neural	Latency
Writing (n=22)	100%	86.4%	p50 181 ms
Computer (n=34)	100%	—	mostly fast-path
Multi-step (n=21)	100%	—	mostly fast-path
Overall (n=56)	100%	—	—
IR honesty (n=112, no fast paths)	—	46.4%	—

Also: user-demand writing honesty (v0.3 lineage) ~80% neural / 100% shipped with structural polish.

Bundle layout

writer/                 # optional promoted_core_100.safetensors seed adapter
writer_unified/         # production writer LoRA (adapters.safetensors)
ir/                     # production IR LoRA (adapters.safetensors)
router.json
manifest.json
inference_config.json

Bases are not vendored here (size). Pull 4-bit bases from MLX community:

• Writer base: mlx-community/LFM2.5-1.2B-Instruct-4bit
• IR base: mlx-community/functiongemma-270m-it-4bit

Hot package class with both 4-bit bases: ~850 MiB primary / 900 MiB hard.

Integration

# pip install mlx-lm huggingface_hub
from huggingface_hub import snapshot_download

bundle = snapshot_download("nsalerni/loudink-v1")
# runner_kind: loudink_v1
# polish_mode: structural
# writer_unified_adapter_path: {bundle}/writer_unified
# ir_adapter_path: {bundle}/ir

See inference_config.json and client_contract.md for Flowcast client fields.

Methods (v0.4)

• Teacher SFT → RFT / expert iteration → daily-mac RFT + residual micro gold
• Structural polish (safety net; neural is promotion currency)
• IR honesty micro-SFT from transcript-first gold (sequence-aware)
• Transcript-first multi-step, Finder paths, Gmail compose, Shortcuts, web search

License

Apache-2.0 (adapters). Base model licenses apply to the community 4-bit bases.