---
license: apache-2.0
language:
- en
pipeline_tag: text-generation
base_model:
- Surpem/Supertron2.1-0.6B
tags:
- gguf
- llama-cpp
- qwen
- qwen3
- chat
- quantized
- q4
- q8
- f16
- local-llm
- surpem
---

# Supertron2.1-0.6B-GGUF

**Supertron2.1-0.6B-GGUF** contains GGUF exports of **Surpem/Supertron2.1-0.6B**, a compact Qwen3-based generalist model by **Surpem**.

This repository is for local inference with `llama.cpp`, LM Studio, Jan, KoboldCpp, text-generation-webui, and other GGUF-compatible runtimes. The original Transformers checkpoint is available at `Surpem/Supertron2.1-0.6B`.

## Available Files

| File | Type | Size | Recommended Use |
| :-- | :-- | --: | :-- |
| `gguf/Supertron2.1-0.6B-F16.gguf` | F16 | ~448 MiB | Highest quality GGUF, larger memory use |
| `gguf/Supertron2.1-0.6B-Q8_0.gguf` | 8-bit | ~610 MiB | Strong quality, efficient local use |
| `gguf/Supertron2.1-0.6B-Q4_K_M.gguf` | 4-bit K-quants | ~378 MiB | Small, fast, best for low-memory devices |

## Which GGUF Should I Use?

### Q4_K_M

Use this when you want the smallest practical model.

Good for:

* laptops
* CPU inference
* fast testing
* low VRAM
* general chat

Tradeoff: slightly lower quality than Q8/F16.

### Q8_0

Use this when you want better quality while keeping the file smaller than full precision.

Good for:

* local coding help
* math prompts
* better instruction following
* GPU offload with modest VRAM

Tradeoff: larger than Q4.

### F16

Use this when quality matters most and memory is available.

Good for:

* comparison testing
* re-quantization
* quality checks
* development workflows

Tradeoff: largest runtime memory use.

## llama.cpp Usage

Install or build `llama.cpp`, then run:

```bash
llama-cli \
  -m gguf/Supertron2.1-0.6B-Q4_K_M.gguf \
  -p "Write a Python function that returns the nth Fibonacci number." \
  -n 256
```

For chat-style prompting:

```bash
llama-cli \
  -m gguf/Supertron2.1-0.6B-Q8_0.gguf \
  -cnv \
  --color \
  -p "You are Supertron, a helpful coding and math assistant."
```

With GPU offload:

```bash
llama-cli \
  -m gguf/Supertron2.1-0.6B-Q4_K_M.gguf \
  -ngl 99 \
  -p "Explain binary search in simple terms." \
  -n 300
```

## llama-server

```bash
llama-server \
  -m gguf/Supertron2.1-0.6B-Q4_K_M.gguf \
  -c 4096 \
  -ngl 99 \
  --host 0.0.0.0 \
  --port 8080
```

Then call it with an OpenAI-compatible client.

## Ollama Modelfile

Create a file named `Modelfile`:

```text
FROM ./gguf/Supertron2.1-0.6B-Q4_K_M.gguf

PARAMETER temperature 0.7
PARAMETER top_p 0.8
PARAMETER top_k 20
PARAMETER num_ctx 4096

SYSTEM """
You are Supertron, a helpful assistant focused on math, coding, and general knowledge.
"""
```

Create and run:

```bash
ollama create supertron2.1-0.6b -f Modelfile
ollama run supertron2.1-0.6b
```

## Recommended Settings

For coding and math:

```text
temperature: 0.2
top_p: 0.8
top_k: 20
repeat_penalty: 1.05
```

For chat:

```text
temperature: 0.7
top_p: 0.8
top_k: 20
repeat_penalty: 1.05
```

For deterministic answers:

```text
temperature: 0.0
```

## Model Line

* Original model: `Surpem/Supertron2.1-0.6B`
* GGUF model: `Surpem/Supertron2.1-0.6B-GGUF`
* MLX 4-bit: `Surpem/Supertron2.1-0.6B-MLX-4Bit`
* MLX 8-bit: `Surpem/Supertron2.1-0.6B-MLX-8Bit`

## Notes

The GGUF files were converted from the latest Supertron2.1-0.6B Transformers checkpoint using llama.cpp tooling. Quantized models are approximations of the original bf16 checkpoint, and behavior can vary by runtime, prompt format, and sampling settings.

## Limitations

* Q4 is smaller but less precise than Q8/F16.
* The model can hallucinate or produce wrong code.
* Human review is recommended for math, code, and factual claims.
* Do not use this model for safety-critical decisions.

## License

Apache 2.0.