dale-0.8b-GGUF

dale-0.8b-GGUF

A small 4-bit GGUF model that turns a plain request in Spanish or English into one Linux bash command. It runs offline on a CPU inside the dale tool. For people who want help writing shell commands.

Made by manuelgscom on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: dale-0.8b is a small language model, fully fine-tuned from Qwen3.5-0.8B, that answers a request in Spanish or English with one bash command for Linux. The answer comes back as a small JSON object. This repository holds its GGUF file in Q4_K_M, made for llama.cpp. The dale tool uses it to turn requests into commands, working offline on a CPU.

Who it suits: It suits people who work in a Linux shell and want a quick command written from a plain sentence, without sending the request to a cloud service.

Industry | Who uses it | How they use it | Example
Software | Developers | Turn a plain request into a bash command | Ask for a command that finds the biggest files
Operations | System admins | Get a command plus a risk rating | Check an unfamiliar command before running it
Education | Learners | Learn bash from explained commands | Ask in Spanish and read the short explanation
Privacy-minded teams | Staff on closed networks | Run the model offline on a CPU | A machine that has no internet access

If you want commands for Windows PowerShell, you can skip it, since the card targets bash on Linux.

Where to use it:

• Linux shell sessions, where the commands will run
• The dale command line tool, offline and on a CPU
• llama.cpp or llama-server, which load the GGUF file directly
• Windows, macOS or Linux machines for downloading the files

Use case | Who | What they get back
Command lookup | Developers | One bash command inside a small JSON object
Learning bash | Learners | A short explanation of the command
Safety check | System admins | A risk rating from dale's own fixed checker

How to set it up: You need these first:

• Python, for the Hugging Face download command
• A CPU machine, since the model is small and runs without a GPU
• Disk space for the file, about 529.3 MB for Q4_K_M
• The dale tool, whose source the card links, or llama.cpp to load the file yourself

1. Install the Hugging Face hub tool.
2. Download the GGUF file.
3. Use it through dale, or start it in llama-server and send requests to it.

pip install -U huggingface_hub
hf download manuelgscom/dale-0.8b-GGUF

How to use it: Type your request in Spanish or English. The tool proposes a command, explains it, shows a risk rating and waits for your confirmation before it runs anything. Read every command first, because the model can be wrong.

Tool | What it does
dale | The offline command line tool that runs this model on a CPU
llama-server | Serves the GGUF file, which the card used for its tests
Verifier | A fixed checker in dale that rates the risk, not the model
Q4_K_M file | The 4-bit quantized GGUF in this repository

Example: You ask, "Find the largest files in this folder." The model answers with one bash command in a small JSON object. The tool shows the command with a short explanation and a risk rating, and it runs only after you confirm.

For newcomers and experts: A GGUF file is a model packed to run on your own computer. A good first step is to try a few harmless requests and read the commands before you run them.

The card describes a full fine-tune, so all the weights were trained on the task. It reports 79.4 % on its own public test of 471 cases, against 62.8 % for larger models without fine-tuning. These are the author's own numbers. The model does not judge danger. A fixed checker does, and dangerous commands need a confirmation word.

Type a request in Spanish or English
The model writes one bash command
The checker rates the risk
You confirm and the command runs

Good to know: Larger siblings score higher on the same test, but they need more memory and some need a GPU.

Title: Accuracy on the author's test v2 in percent
dale-0.8b: 79.4%
dale-2b: 81.7%
dale-4b: 87.0%
dale-9b: 89.2%
Qwen3.5-9B without fine-tuning: 62.8%

• Mistakes: the card says about 1 case in 5 is wrong on its test, so check each command
• Speed: the card reports a median reply time of 1.72 s on an ordinary 2020 laptop CPU, with no GPU
• Family: dale-2b, dale-4b and dale-9b are larger models in the same line
• License: Apache 2.0
• Cost to run: none named, you supply the machine
• Data it reaches: the requests you type, which stay on your machine when it runs offline

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 752M
• Weights: about 376 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 2,163
• Likes: 44
• License: apache-2.0
• Languages: es, en
• Based on: manuelgscom/dale-0.8b
• Tags: gguf, bash, shell, nl2bash, llama.cpp, text-generation, model-index, imatrix, conversational

Links

• Model: https://huggingface.co/manuelgscom/dale-0.8b-GGUF
• Files: https://huggingface.co/manuelgscom/dale-0.8b-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download manuelgscom/dale-0.8b-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.