KiwiMate/ KiwiMate-Small-1.0:F16

7 1 month ago

The first (non-preview) version of KiwiMate Small

vision 
ollama run KiwiMate/KiwiMate-Small-1.0:F16

Details

1 month ago

26de24d6be74 · 19GB

model
qwen35
·
9.2B
·
F16
projector
clip
·
456M
·
BF16
template
{{- if .System -}} <|im_start|>system {{ .System }}<|im_end|> {{- end -}} {{- range $i, $_ := .Messa
system
You are KiwiMate a friendly ai.
params
{ "temperature": 0.0001 }

Readme

license: apache-2.0 base_model: Qwen/Qwen3.5-9B language: - en - mi tags: - gguf - llama.cpp - unsloth - vision-language-model - conversational - endpoints_compatible - new-zealand pipeline_tag: image-text-to-text

library_name: llama.cpp

KiwiMate-Small-1.0

KiwiMate is an AI companion app for iOS and Android with a distinctly New Zealand cultural identity — everyday NZ English, te reo Māori, and Kiwi context baked into how it talks. KiwiMate-Small-1.0 is the mid-tier model in the KiwiMate family, sitting between Mini and Medium for everyday chat with vision support.

Model Details
Base model Qwen/Qwen3.5-9B
Parameters ~9B
Architecture Hybrid linear/full attention (24 linear-attention layers, 8 full-attention layers, 32 total)
Context length 262,144 tokens
Modality Text + image (vision-language)
Fine-tuning Unsloth, 2x faster training
Format GGUF (for llama.cpp)
License Apache 2.0

KiwiMate-Small-1.0 was fine-tuned from Qwen3.5-9B on KiwiMate’s proprietary dataset of New Zealand English, te reo Māori, and Kiwi cultural context, tuned for the AI-companion persona used across the KiwiMate app.

Licensing note: Qwen3.5-9B is released under Apache 2.0, which permits commercial use, fine-tuning, and redistribution without royalties. This is a cleaner license position than KiwiMate Mini’s Qwen2.5-VL-3B base, which ships under the more restrictive Qwen Research license.

Available Files

File Type Size Use case
KiwiMate-Small-1.0.F16.gguf Full precision 18.4 GB Evaluation, highest fidelity
KiwiMate-Small-1.0.Q8_0.gguf 8-bit quant 9.8 GB Near-lossless, lower VRAM
KiwiMate-Small-1.0.Q4_K_M.gguf 4-bit quant 5.8 GB Best for constrained hardware / edge deployment
KiwiMate-Small-1.0.BF16-mmproj.gguf Vision projector 921.7 MB Required alongside any of the above for image input

Usage

llama.cpp

Text-only:

llama-cli -hf KiwiMate/KiwiMate-Small-1.0 --jinja

Multimodal (text + image):

llama-mtmd-cli -hf KiwiMate/KiwiMate-Small-1.0 --jinja

Serving as an OpenAI-compatible endpoint

This model is served in production via a Hugging Face Inference Endpoint backed by llama.cpp. If you’re self-hosting, llama-server exposes the same OpenAI-compatible API:

llama-server -hf KiwiMate/KiwiMate-Small-1.0 --jinja --port 8080

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "Kia ora, what'\''s a good spot for fish and chips in Christchurch?"}]
  }'

Note: if you’re hitting an HF Inference Endpoint Playground and getting unexplained 404s, they’re usually a request-routing issue during container restarts — calling /v1/chat/completions directly tends to work even when the Playground doesn’t.

Intended Use

Casual conversation, everyday NZ-context Q&A, and lightweight image understanding within the KiwiMate app. For heavier reasoning tasks, route to KiwiMate-Medium-1.0 or KiwiMate-Large-1.0.

Limitations

  • Fine-tuned for a specific companion persona; not intended as a general-purpose assistant.
  • Vision understanding inherits the base model’s ViT encoder; expect reduced accuracy on dense text-in-image (OCR-heavy) tasks.
  • As with any fine-tune, verify outputs before using in production-critical contexts.

Acknowledgements

Fine-tuned and converted to GGUF using Unsloth.