KiwiMate-Small-Preview
KiwiMate-Small-Preview is the larger, multimodal model in the KiwiMate AI companion app's model family, fine-tuned for richer conversation and image understanding on the app's higher subscription tiers.
⚠️ Preview status: This is a prototype release. The name reflects its preview status — expect breaking changes, retraining, and behavioural shifts before a stable v1 release.
Model Details
| Developed by | KiwiMate / KyleCodeKiwi |
| Base model | Qwen3.5-9B |
| Architecture | Qwen3.5 (qwen3_5) |
| Parameters | ~9.65B |
| Modality | Multimodal (image-text-to-text) |
| Fine-tuning framework | Unsloth |
| License | Apache 2.0 |
| Languages | English (with New Zealand English and Te Reo Māori vocabulary coverage) |
| Model class | AutoModelForMultimodalLM |
| Inference compatibility | vLLM-compatible |
Intended Use
KiwiMate-Small-Preview is intended for the higher-tier, "smarter" conversational experience inside the KiwiMate app, providing:
- Deeper general-purpose chat and assistant-style conversation than KiwiMate-Mini-Preview
- Image understanding (e.g. user-uploaded photos within the app)
- New Zealand cultural and "Kiwi" context awareness, plus light Te Reo Māori vocabulary
- KiwiMate app-specific lore and identity content
It is not intended for high-stakes, medical, legal, or financial advice, and should not be treated as authoritative on Te Reo Māori or tikanga matters.
Training Data
Fine-tuned using the same category structure as the rest of the KiwiMate model family: NZ English, Te Reo Māori, KiwiMate Origin (app-specific lore/identity), NZ Fun Facts, and MiniGame Knowledge, alongside the base Qwen3.5-9B multimodal capabilities.
Files & Quantizations
Distributed as safetensors (full precision) and GGUF quantizations for flexible serving:
| Format | Use case |
|---|---|
| F16 | Highest fidelity, largest size |
| Q6_K | Near-lossless, smaller footprint |
| Q4_K_M | Balanced quality/size |
| Q2_K_L | Smallest footprint, lowest fidelity |
Deployment
Served in production via a Hugging Face Inference Endpoint on an L40S GPU, fronted by a Supabase Edge Function (OpenAI-compatible proxy) that routes KiwiMate app traffic to this and other KiwiMate model endpoints behind a single API.
Known Limitations
- An earlier deployment issue caused by a
tokenizer_config.jsonincompatibility (TokenizersBackend→Qwen2TokenizerFast) was identified and resolved; deployments should pin a compatible tokenizer backend. - As with any 9–10B parameter multimodal model, image reasoning and fine factual detail can lag behind larger frontier models — it is tuned for a balance of capability and serving cost rather than maximum raw performance.
License
Released under the Apache 2.0 license, consistent with the open weights commitment for the KiwiMate model family.