800M Parameter Model Beats Qwen 9B and Gemma 4 in Linux Command...

800M Parameter Model Beats Qwen 9B and Gemma 4 in Linux Command Generation

Developer Manuel García Sanz published the dale model family on Hugging Face—fine-tuned versions of Qwen3.5 that translate English or Spanish requests into a single Linux bash command. dale-0.8b scores 79.4% on a 471-task test, compared to 62.8% for base Qwen3.5-9B and Gemma 4 E4B-it.

Compiled by Sergey KostenchukPublished 2026-11-09Updated 2026-11-10

2026-11-09 Coding HuggingFace

🛠 800M Parameter Model Beats Qwen 9B and Gemma 4 in Linux Command Generation

Developer Manuel García Sanz released the dale models on Hugging Face—fine-tuned versions of Qwen3.5 that translate English or Spanish requests into a single Linux bash command with an explanation. On a 471-task test, dale-0.8b scores 79.4% compared to 62.8% for base Qwen3.5-9B and Gemma 4 E4B-it.

🌍 This confirms the trend toward cheap local models for narrow tasks: dale-0.8b is 11 times smaller than the 9B model and runs on CPU, while a deterministic C++ verifier, not the model, determines the risk level of the command.

👤 It works offline: the binary is ~51 MB (MIT), and the GGUF version takes up 529 MB, responding in a median of 1.7 seconds on a 2020 laptop. The model makes a mistake in 1 out of 5 cases—the command should be re-read; on GPUs with 4+ GB, dale-4b (87.0%) and dale-9b (89.2%) are available.

Source 1: https://huggingface.co/manuelgscom/dale-0.8b-GGUF Source 2: https://github.com/manuelgscom/dale

Sources