KiwiMate Mini is the lightweight, on-device model tier for KiwiMate, an AI companion app for iOS and Android with a New Zealand cultural identity. Mini 1.1 is fine-tuned from Qwen3-VL-2B-Thinking, replacing Mini 1.0’s Qwen2.5-VL-3B-Instruct base to add reasoning (“thinking”) capabilities and move the tier onto an Apache 2.0 base.

This model is intended for fast, low-resource inference — locally on-device or via lightweight self-hosted endpoints — while KiwiMate Small and Medium handle heavier multimodal tasks server-side.

⚠️ Preview build. This is an early release of the Mini 1.1 line for internal testing ahead of KiwiMate app integration. Expect rough edges, especially around reasoning-trace formatting and long conversations.

What’s new vs. Mini 1.0 (Qwen2.5-VL-3B)

License fixed — Mini 1.0 inherited Qwen2.5-VL-3B's Qwen Research license (non-commercial). Mini 1.1's base, Qwen3-VL-2B-Thinking, is Apache 2.0, clearing the licensing concern flagged on the 1.0 line.
Reasoning support — this is a "Thinking" checkpoint: it emits a <think>...</think> reasoning trace before its final answer, rather than answering directly. Any client or server-side proxy that displays or forwards Mini's output should account for this (strip or surface the trace deliberately — don't pass it straight to the end user unfiltered).
Smaller base (2B vs. 3B params), though real-world speed/quality will depend on which quant you run.
New base architecture tag: qwen3_vl (Qwen3-VL) instead of qwen2_5_vl.

Files File Size Description KiwiMate-Mini-1.1.F16.gguf 3.4 GB Full-precision (FP16) weights, highest quality, largest footprint KiwiMate-Mini-1.1.BF16.gguf 3.4 GB Full-precision (BF16) weights, alternative to F16 KiwiMate-Mini-1.1.Q8_0.gguf 1.8 GB Near-lossless quantization KiwiMate-Mini-1.1.Q6_K.gguf 1.4 GB Good balance of quality/size KiwiMate-Mini-1.1.Q4_K_M.gguf 1.1 GB Recommended for constrained/on-device use KiwiMate-Mini-1.1.IQ4_XS.gguf 1.0 GB Smallest, fastest, largest quality tradeoff KiwiMate-Mini-1.1.BF16-mmproj.gguf 822 MB Vision projector — required alongside a text GGUF above for image input

As with the Mini 1.0 line, spot-check quality at Q4_K_M / IQ4_XS relative to F16/Q6_K once you’ve run your usual eval pass — the 2B Thinking base is architecturally different from 1.0’s 3B base, so the quantization tradeoff may not track identically. Usage llama.cpp

Vision input requires loading both the text model and the mmproj file:

llama-mtmd-cli
-hf KiwiMate/KiwiMate-Mini-1.1
–mmproj KiwiMate-Mini-1.1.BF16-mmproj.gguf
-p “Describe this image.”
–jinja

Text-only:

llama-cli -hf KiwiMate/KiwiMate-Mini-1.1 –jinja

Ollama

Ollama can pull GGUF files directly from a Hugging Face repo without a local Modelfile:

ollama run hf.co/KiwiMate/KiwiMate-Mini-1.1:Q4_K_M

Chat template

Qwen3 ChatML format, –jinja handles this automatically with the commands above:

<|im_start|>system {system_prompt}<|im_end|> <|im_start|>user {user_message}<|im_end|> <|im_start|>assistant

Because this is a Thinking checkpoint, the assistant turn will typically begin with a … block before the actual reply. If you’re serving this behind a proxy (e.g. runpod-chat/index.ts / hf-chat), make sure your response handling parses or strips the think block rather than forwarding it verbatim, and update stop-sequence config if you’re migrating directly from a non-thinking base. Training

Base model: Qwen/Qwen3-VL-2B-Thinking
Framework: Unsloth (FastVisionModel)
Quantization: GGUF via llama.cpp (F16, BF16, Q8_0, Q6_K, Q4_K_M, IQ4_XS)
Fine-tuning focus: New Zealand English, Te Reo Māori awareness and cultural sensitivity, KiwiMate persona and identity, conversational NZ context, and stabilising reasoning-trace behaviour on a smaller base

Intended use

Conversational AI companion for the KiwiMate app — casual chat, light image understanding (e.g. describing a photo, reading a screenshot), and NZ-flavoured personality, now with lightweight step-by-step reasoning for questions that benefit from it. Not intended for high-stakes reasoning, factual research, or professional/medical/legal advice. License

This model is derived from Qwen3-VL-2B-Thinking, distributed by Qwen under the Apache 2.0 license — unlike Mini 1.0, which carried Qwen2.5-VL-3B-Instruct’s non-commercial Qwen Research license. Review the base model’s license before redistributing, but this tier no longer carries the commercial-use restriction that applied to Mini 1.0. Links

KiwiMate app: kiwimate.net
Dev logs: YouTube — KyleCodeKiwi