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. The smallest model, dale-0.8b, scored 79.4% on the public dale-bash-test-v2 test of 471 tasks, while full-size Qwen3.5-9B and Gemma 4 E4B-it without fine-tuning achieve 62.8%. The model works fully offline on CPU via llama.cpp, a deterministic code evaluates the risk of the proposed command rather than a neural network, and execution occurs only after user confirmation.

image

What happened

Manuel García Sanz released four models on Hugging Face — dale-0.8b, dale-2b, dale-4b, and dale-9b — fine-tuned versions of Qwen3.5 that accept English or Spanish requests and return JSON with fields type (command, refuse, unsure), command, and explanation: the model either proposes a command, refuses, or acknowledges uncertainty. On the same public dale-bash-test-v2 test of 471 tasks, unfine-tuned Qwen3.5-4B scores 54.6%, the base Qwen3.5-0.8B of the same size as the smallest dale — 14.9%, while the larger versions of the family reach 87.0% (dale-4b) and 89.2% (dale-9b). The dale-0.8b model in GGUF Q4_K_M format weighs 529.3 MB and runs on CPU via llama.cpp on a 2020 laptop with four cores and eight threads with a median latency of 1.72 seconds, and its weights are distributed under the Apache-2.0 license. Along with the models, a CLI tool dale under the MIT license was published — a single native binary of about 51 MB, which marks the proposed command with risk labels safe/caution/danger and requests confirmation before execution.

Context

The task of translating natural language into shell commands (nl2bash) requires the model to convert a human formulation into an executable command without errors in flags and syntax, and universal LLMs stumble over strict output format requirements. The main resource in dale is not size, but narrow fine-tuning: it is the quality of the training dataset that, judging by the gap from the base model of the same size, played a decisive role in the accuracy increase. The effect of "specialization beats scale" in narrow domains is a known and repeatedly described phenomenon in the industry, so the result demonstrates an established SFT practice, not a new type of architecture, although within the family, scale is still useful and larger versions are noticeably more accurate than smaller ones. Architecturally, the project implements a hybrid "neural network proposes, deterministic code decides" template: the model only returns a command variant with an explanation, while the assessment of danger and permission to execute remain with regular code — a common approach to the security of agents working with the terminal. The case continues a broader industry movement toward small local specialized models as an alternative to expensive frontier calls to large models.

Why this matters for the industry

For AI product creators, dale is interesting primarily as a reproducible template for a cheap local AI function: a fine-tuned sub-gigabyte model returns structured JSON with three states (command, refuse, unsure), and the danger of the command is evaluated by a deterministic verifier, not the model itself, and such a scheme can be transferred to other narrow domains. An offline nl2bash function can be embedded into a product for Linux x86_64 today: weights under Apache-2.0 are distributed for free and installed with one command, which means developers of wrappers over large APIs in this niche get their first cheap competitor. The case confirms the trend where access to a large model ceases to be an advantage in narrow structured tasks: a specialized model about 11 times smaller than its competitor solves them better with almost no infrastructure costs. For production embedding, data is currently lacking: dale has no API and serving stack, and the throughput of the solution has not been measured.

Why this matters for users

dale is ready for use on a personal machine right now and works fully offline, so request formulations do not leave the user's device. Installation is reduced to one command: a native binary of about 51 MB requires Linux x86_64 with AVX2 support and glibc 2.29+. A typical call looks like dale "find the files over 200 MB in this folder": the model proposes a command, briefly explains what it does, marks the danger level as safe/caution/danger, and asks for confirmation before execution. Approximately one command in five in the test turns out to be incorrect, so the ready text should be reread with your own eyes before execution. If there is a GPU with 4 GB VRAM or more, it makes sense to switch to the larger dale-4b and dale-9b, which are significantly more accurate. The boundaries are strict: only bash and only Linux are supported, without zsh, fish, and macOS.

What is still unknown / limitations

Almost all published characteristics are author measurements: the count is conducted on the author's own benchmark dale-bash-test-v2, confidence intervals are not provided, and the configuration of baselines, including zero-shot or few-shot prompting, output format, and temperature, is not described, so a gap of about 78 tasks out of 471 without independent verification may turn out to be fragile. The test is public and verifiable by third parties, but the intersection of its distribution with training data cannot be ruled out — a typical risk of an author's benchmark. The composition, volume, and origin of the fine-tuning dataset have not been disclosed, although it is the quality of this dataset that most likely ensured the main jump relative to the base model of the same size. Predictions about the mass copying of the "three states plus deterministic verifier" template and about micro-models for other command languages are interpretation, not fact from sources.

Sources

Author

Look at AI, editorial team