flowcast-v5-5-lite v0.2.0
Neural-first sub-600MB stack: unified writer LoRA + three-tier polish (minimal → selective → structural).
Predecessor: nsalerni/flowcast-v5-lite
Production stack
- Writer: single unified LoRA on LFM2.5-1.2B (no hybrid paraphrase/longform adapters)
- Polish: three-tier deterministic rescue (structural tier for longform/email/SQL)
- IR: FunctionGemma IR v5 + learned transcript router
User-demand benchmarks (shipped)
| Suite | Cases | Target |
|---|---|---|
| Computer | 126 | 100% |
| Writing | 81 | 100% |
| Paraphrase | 15 | ≥70% |
| Combined (computer+writing) | 207 | ≥95% |
See inference_config.json and manifest.json for integration fields.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support