0.8B-Parameter Model Beats Qwen 9B at Generating Bash Commands: 79% vs 63%

Anyone who works in a terminal knows the feeling: you know exactly that you need to find large files, but you can't remember which flags find takes or how to sort the output properly. Usually people turn to a search engine or open a chat with a large model. The first is slow; the second requires internet access, an API key, and sending data out. The author of the dale project took a different approach: they built a small model that lives right on your machine and does just one thing.

What dale is

dale-0.8b-GGUF has appeared on Hugging Face, a specialized model with 0.8 billion parameters. You type a request in plain language, for example, "Find all large files in the current directory." In response, the model returns a single Bash command in structured JSON and a short explanation of what that command does.

Two things set it apart from a "mere command generator." First, the model can decline: if a request is not suitable, it answers refuse. Second, if a request is too vague, it doesn't guess but returns unsure. For a tool that can potentially run commands on your system, this matters more than a few extra percentage points of accuracy: an honest "not sure" is better than a confident but destructive command.

The numbers and what is behind them

The author tested the models on a set of 471 tasks. The results are as follows:

• dale-0.8B — 79.4%;
• Qwen3.5-9B — 62.8%;
• Gemma 4 E4B — 62.8%;
• Qwen3.5-4B — 54.6%;
• Qwen3.5-0.8B — 14.9%.

The last line is the most telling. A general-purpose model of the same size barely copes with the task, while the specialized one outperforms models 5–10 times larger. This is further confirmation of a long-familiar idea: on a narrow, well-defined task with a strict output format, a small fine-tuned model can beat generalists. It doesn't need to know everything in the world; it is enough for it to know Bash and the output format well.

The hardware requirements are modest: Q4_K_M quantization, a file of about 529 MB, running on a CPU via llama.cpp. Median latency is 1.72 seconds on a 2020 laptop. No GPU is needed, and no network either. The license is Apache 2.0, meaning the model can be used in commercial projects too.

A CLI tool, also called dale, is built on top of it. It suggests a command, asks for confirmation before executing it, and classifies dangerous operations. Such a layer of protection is sensible: a model with 0.8 billion parameters still gets it wrong in roughly one case out of five.

Limitations

The main caveat: all the figures are reported by the author themselves. We don't know how similar the test set is to the training data or how it was put together. It's possible the tasks were chosen to suit the model's strengths. There are no independent measurements yet, and 79.4% is still not "almost always correct." Moreover, the test covers Bash on Linux: nothing is known about macOS specifics, PowerShell, or exotic utilities.

What to do with this

Developers should download the model and run it on their own real queries: your own sample will tell you more than someone else's 471 tasks. Teams building internal tools will find the pattern itself useful: a small local model, structured JSON, explicit refuse and unsure branches, and human confirmation before execution. For those who care about privacy or work without internet, this is a ready-made foundation that doesn't need to be connected to a cloud API.