You can deploy this model and start earning money today!

Quatfit/
Quatfit-Mini-GGUF

Name: Quatfit-Mini-GGUF
Availability: InStock

Views

No views yet

Quatfit Mini — GGUF

Every llama.cpp Quantization of Quatfit Mini 8B · 131K Context · Consumer-Hardware Ready

Overview

This repository provides GGUF-quantized builds of Quatfit Mini, an 8B-parameter multimodal model built on the Gemma 4 architecture. Every legacy and K-quant format offered by llama.cpp — from F16 down to the smallest 2-bit Q2_K — is provided here, converted directly from the published FP32 base weights. Each text quant ships together with its matching mmproj file, so vision and audio input work out of the box with every download.

Use this repository if you want to run Quatfit Mini locally with llama.cpp, Ollama, LM Studio, Jan, Open WebUI, koboldcpp, or any other GGUF-compatible runtime.

For fine-tuning, research, or maximum numerical fidelity, use the base FP32 repository instead: Quatfit/Quatfit-Mini.

Available Quantizations

All sizes below are approximate (computed from the 8B parameter count and each format's bits-per-weight; actual file size varies slightly by build). "Quality" is a rough relative ranking, not a benchmark score — see Choosing a Quant below.

K-Quants and Legacy Quants

File	Bits/Weight	Approx. Size	Quality	Notes
`Quatfit-Mini-F32.gguf`	32	~32.1 GB	Reference	Unquantized, identical to source weights
`Quatfit-Mini-F16.gguf`	16	~16.1 GB	Reference	Half precision, negligible loss vs FP32
`Quatfit-Mini-Q8_0.gguf`	8.5	~8.5 GB	Excellent	Near-lossless, largest practical quant
`Quatfit-Mini-Q6_K.gguf`	6.6	~6.6 GB	Excellent	Very close to F16 quality
`Quatfit-Mini-Q5_K_M.gguf`	5.7	~5.7 GB	Very Good	Recommended high-quality default
`Quatfit-Mini-Q5_K_S.gguf`	5.5	~5.6 GB	Very Good	Slightly smaller than Q5_K_M
`Quatfit-Mini-Q5_1.gguf`	6.0	~6.1 GB	Good	Legacy format
`Quatfit-Mini-Q5_0.gguf`	5.5	~5.6 GB	Good	Legacy format
`Quatfit-Mini-Q4_K_M.gguf`	4.8	~4.9 GB	Good	Most popular general-purpose quant
`Quatfit-Mini-Q4_K_S.gguf`	4.6	~4.7 GB	Good	Smaller, slightly lower quality than Q4_K_M
`Quatfit-Mini-Q4_1.gguf`	5.0	~5.1 GB	Good	Legacy format
`Quatfit-Mini-Q4_0.gguf`	4.5	~4.7 GB	Good	Legacy format, wide compatibility
`Quatfit-Mini-Q3_K_L.gguf`	4.3	~4.3 GB	Moderate	Largest 3-bit K-quant
`Quatfit-Mini-Q3_K_M.gguf`	3.9	~4.0 GB	Moderate	Balanced 3-bit option
`Quatfit-Mini-Q3_K_S.gguf`	3.5	~3.7 GB	Lower	Smallest 3-bit K-quant
`Quatfit-Mini-Q2_K.gguf`	2.6	~3.2 GB	Low	Smallest legacy/K-quant, noticeable degradation

Choosing a Quant

You have ≥ 24 GB VRAM (RTX 3090/4090, A6000): Q6_K or Q8_0 for near-reference quality.
You have 8–12 GB VRAM (RTX 3060/3070/4060 Ti): Q4_K_M is the standard recommendation — the best balance of size, speed, and quality.
You have 6–8 GB VRAM: Q4_K_S or Q4_0.
You have < 6 GB VRAM or are running on CPU only: Q3_K_M or Q2_K, with an expected drop in reasoning/coding accuracy.
You need maximum fidelity for evaluation or further conversion: F16 or F32.

As a rule of thumb, prefer the K-quants (Q_K_M variants) for general use — they offer the best quality-per-byte among the legacy and K-quant families. The plain legacy formats (Q4_0, Q4_1, Q5_0, Q5_1) remain useful mainly for compatibility with older runtimes.

Usage

llama.cpp (CLI)

bash

./llama-cli \
  -hf Quatfit/Quatfit-Mini-GGUF:Q4_K_M \
  -p "Explain the difference between GQA and MHA." \
  -n 512

llama.cpp (server)

bash

./llama-server \
  -hf Quatfit/Quatfit-Mini-GGUF:Q4_K_M \
  --host 0.0.0.0 \
  --port 8080 \
  -c 131072

Manual download

bash

huggingface-cli download Quatfit/Quatfit-Mini-GGUF \
  Quatfit-Mini-Q4_K_M.gguf \
  --local-dir .

./llama-cli -m Quatfit-Mini-Q4_K_M.gguf -p "Hello!" -n 256

Ollama

ollama run hf.co/Quatfit/Quatfit-Mini-GGUF:Q4_K_M

LM Studio / Jan / Open WebUI

Search for Quatfit/Quatfit-Mini-GGUF directly in-app and select your preferred quant from the list above.

Prompt Format

Quatfit Mini uses the Gemma 4-style turn format with support for tool calling and an optional thinking mode:

<|turn>user
Your message here<turn|>
<|turn>model

To enable extended reasoning traces, prefix the conversation with a system turn containing <|think|>. The model will emit its reasoning inside <|channel>thought ... <channel|> before the final response.

Most GGUF-compatible chat UIs (Ollama, LM Studio, Open WebUI, llama-server's built-in UI) apply this template automatically via the embedded chat template in each .gguf file.

Multimodality Note

These GGUF conversions currently expose the text capabilities of Quatfit Mini. Vision and audio inputs require the multimodal projector files (mmproj.gguf), which are provided alongside the text-only quants for runtimes that support llama.cpp multimodal (llama-mtmd-cli, compatible server builds). Consult your runtime's documentation for multimodal GGUF support before relying on image/audio input outside of the base FP32 Transformers checkpoint.

Relationship to the Base Model

	Quatfit Mini (base)	Quatfit Mini GGUF (this repo)
Format	`safetensors`, FP32	`.gguf`, quantized (F32 down to Q2_K)
Best for	Fine-tuning, research, max fidelity	Local inference, consumer hardware
Runtime	🤗 Transformers	llama.cpp and compatible runtimes
Size	~32 GB	~2 GB – ~32 GB depending on quant

Full architecture details, training recipe, and benchmark methodology are documented in the Quatfit Mini Technical Report.

Responsible AI

Quatfit Mini may generate inaccurate, biased, or inappropriate outputs, and quantization can mildly amplify errors at lower bit-widths (particularly Q2_K and Q3_K_S). For production deployments:

Verify critical information
Apply RAG for factual grounding
Use application-level safety filters
Prefer Q4_K_M or higher for anything beyond casual/experimental use
Keep human oversight for high-risk domains

Citation

bibtex

@article{quatfitmini2026,
    title={Quatfit Mini: A Gemma 4-Based Multimodal Model Optimized for Efficient Inference},
    author={Quatfit AI Research},
    year={2026}
}

License

Apache License 2.0. See LICENSE for details.

🤗 GGUF Repository • 🧠 Base Model (FP32) • 📄 Technical Report