π Strawberry
Collection
A collection of multilingual reasoning models named Strawberry β’ 7 items β’ Updated β’ 1
π Strawberry GGUF
Strawberry GGUF is a collection of quantized GGUF versions of artindnr/strawberry, a fine-tuned version of openai/gpt-oss-20b specialized for Farsi (Persian) reasoning traces and multilingual chain-of-thought generation.
These files are optimized for efficient local inference using llama.cpp, LM Studio, KoboldCpp, and other GGUF-compatible inference engines.
Model Details
- Model name: Strawberry
- Original base model:
openai/gpt-oss-20b - Fine-tuned model:
artindnr/strawberry-1 - Architecture:
gpt_oss - Parameters: ~21B
- Format: GGUF
- Quantization: Multiple GGUF quantization variants
- Fine-tuned by:
artindnr - License: Apache 2.0
- Languages: Farsi (Persian), English, multilingual reasoning
About Strawberry-1
Strawberry is a fine-tuned version of openai/gpt-oss-20b designed to improve multilingual reasoning capabilities, with a particular focus on native Farsi reasoning traces.
Most reasoning models generate their internal reasoning primarily in English, even when answering in another language. Strawberry-1 was trained to:
- Generate coherent reasoning traces in Farsi
- Reason across multiple languages
- Preserve strong instruction-following ability
- Maintain general reasoning capabilities from the GPT-OSS base model
The model supports both:
- Farsi reasoning
- English reasoning
- Multilingual reasoning workflows
Training
Strawberry was trained using a combination of full fine-tuning and LoRA experiments on top of openai/gpt-oss-20b.
The released model is a merged dense-weight checkpoint before conversion into GGUF format.
Training Data
Strawberry was trained using the Thinking Datasets collection:
artindnr/Persian-Thinking
Farsi reasoning traces.artindnr/Persian-English-Thinking
Mixed Farsi and English reasoning traces.artindnr/Multilingual-Thinking
Multilingual chain-of-thought reasoning.artindnr/Multilingual-Thinking-200
A smaller multilingual reasoning subset.
GGUF Format
GGUF is a format designed for efficient local inference.
These files can be used with:
- llama.cpp
- LM Studio
- KoboldCpp
- text-generation-webui
- Ollama-compatible runtimes
- Other GGUF-supported applications
GGUF provides:
- Reduced memory usage
- Faster local inference
- CPU and GPU acceleration support
- Quantized model execution without requiring the original framework
Available Quantizations
| Quantization | Size |
|---|---|
| Q2_K | 12.1 GB |
| Q3_K_S | 12.1 GB |
| Q3_K_M | 12.9 GB |
| Q3_K_L | 13.3 GB |
| IQ4_XS | 12.2 GB |
| Q4_K_S | 14.7 GB |
| MXFP4_MOE | 12.1 GB |
| Q4_K_M | 15.8 GB |
| Q5_K_S | 15.9 GB |
| Q5_K_M | 16.9 GB |
| Q6_K | 22.2 GB |
| Q8_0 | 22.3 GB |
Recommended Quantizations
β Best Overall Balance
MXFP4_MOE
Recommended for most users.
This quantization is designed specifically for Mixture-of-Experts models and provides a strong balance between:
- Memory usage
- Inference speed
- Model quality
Highest Quality
Q8_0
Q6_K
Recommended when you have enough RAM or VRAM.
These preserve more information from the original model weights and generally provide the closest quality to the original BF16 model.
Smaller Memory Usage
Q2_K
Q3_K_M
IQ4_XS
Useful for:
- Smaller systems
- Limited VRAM environments
- CPU inference
Lower-bit quantizations reduce memory requirements but may slightly reduce reasoning quality.
Usage
llama.cpp
Install llama.cpp:
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
cmake -B build
cmake --build build --config Release
Run Strawberry-1:
./llama-cli \
-m strawberry-1-MXFP4_MOE.gguf \
-p "ΨͺΩ Ϊ©Ϋ ΩΨ³ΨͺΫΨ" \
-n 512
Chat Template
Strawberry uses the GPT-OSS Harmony chat format.
GGUF files contain the required metadata for compatible runtimes to automatically apply the correct chat template.
The model supports reasoning language control through the system prompt.
Example:
reasoning language: Farsi
or:
reasoning language: English
Example Prompt
<system>
reasoning language: Farsi
</system>
<user>
ΪΨ±Ψ§ Ψ’Ψ³Ω
Ψ§Ω Ψ’Ψ¨Ϋ Ψ§Ψ³ΨͺΨ
</user>
The model will generate responses following the GPT-OSS Harmony format, including reasoning and final response channels.
Hardware Requirements
Approximate memory requirements:
| Quantization | Recommended Memory |
|---|---|
| Q2_K | ~16 GB |
| Q3_K_M | ~16 GB |
| IQ4_XS | ~16 GB |
| MXFP4_MOE | ~16 GB |
| Q4_K_M | ~20 GB |
| Q5_K_M | ~24 GB |
| Q6_K | ~28 GB |
| Q8_0 | ~32 GB |
Actual requirements depend on:
- Context length
- Runtime overhead
- GPU offloading settings
- Batch size
Intended Use
Strawberry-1 GGUF is intended for:
- Local AI assistants
- Farsi-language applications
- Multilingual reasoning research
- Offline inference
- Educational tools
- Experimentation with reasoning models
Limitations
- Quantization can slightly reduce reasoning quality compared to the original BF16 model.
- Lower-bit quantizations trade accuracy for reduced memory usage.
- The model may occasionally produce incorrect information or hallucinations.
- Farsi reasoning quality may vary depending on the complexity of the task.
- This model inherits limitations from the original GPT-OSS architecture.
License
Strawberry GGUF is released under the Apache 2.0 license, consistent with the original model license.
Credits
Built on top of:
GGUF conversion and quantization were performed to enable efficient local inference.
Citation
If you use Strawberry in your work, please cite:
@misc{strawberry,
title = {Strawberry: A Farsi and Multilingual Reasoning Model Fine-tuned from GPT-OSS-20B},
author = {artindnr},
year = {2026},
url = {https://huggingface.co/artindnr/strawberry}
}
- Downloads last month
- -
Hardware compatibility
Log In to add your hardware
2-bit
3-bit
4-bit
5-bit
6-bit
8-bit