Details

Updated 3 weeks ago

3 weeks ago

0c029d8b3293 · 18GB

model

archqwen35

parameters27.3B

quantizationQ4_K_M

17GB

projector

archclip

parameters461M

quantizationBF16

931MB

template

{{- if .System -}} system {{ .System }} {{- end -}} {{- range $i, $_ := .Messages }} {{- $last := eq

399B

system

You are KiwiMate a friendly ai.

31B

params

{ "temperature": 0.0001 }

23B

license: apache-2.0 base_model: Qwen/Qwen3.6-27B 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-Medium-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-Medium-1.0 is the upper-mid tier in the KiwiMate model family, built for conversations that need stronger reasoning than Small can offer while staying self-hostable on a single high-VRAM GPU.

Model Details


Base model	Qwen/Qwen3.6-27B
Parameters	~27B
Architecture	Hybrid linear/full attention (48 linear-attention layers, 16 full-attention layers, 64 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-Medium-1.0 was fine-tuned from Qwen3.6-27B 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: Qwen3.6-27B is released under Apache 2.0, which permits commercial use, fine-tuning, and redistribution without royalties — 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-Medium-1.0.BF16-00001-of-00002.gguf` + `...00002-of-00002.gguf`	Full precision (split)	~54.7 GB total	Evaluation, highest fidelity
`KiwiMate-Medium-1.0.Q8_0.gguf`	8-bit quant	29.0 GB	Near-lossless, lower VRAM
`KiwiMate-Medium-1.0.Q4_K_M.gguf`	4-bit quant	16.8 GB	Best for single high-VRAM consumer GPU
`KiwiMate-Medium-1.0.BF16-mmproj.gguf`	Vision projector	931.1 MB	Required alongside any of the above for image input

Usage

llama.cpp

Text-only:

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

Multimodal (text + image):

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

Serving as an OpenAI-compatible endpoint

Intended for deployment behind a Hugging Face Inference Endpoint (llama.cpp backend) or self-hosted with llama-server, which exposes the same OpenAI-compatible API:

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

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "Explain the difference between a tramping track and a walking track in NZ."}]
  }'

Intended Use

Conversations requiring stronger reasoning, longer context, or more nuanced NZ cultural/te reo Māori handling than KiwiMate-Small-1.0. Suited to a single high-VRAM GPU deployment (e.g. L40S-class or better at Q4_K_M/Q8_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.
Larger footprint than Small — not suited to edge/low-VRAM deployment; use Small or Mini for that.
As with any fine-tune, verify outputs before using in production-critical contexts.

Acknowledgements

Fine-tuned and converted to GGUF using Unsloth.

The first version of KiwiMate that has decent intelligence