---
type: "post"
id: "manuel-garcia-s-hugging-face-page-presents-dale-8a0293"
slug: "manuel-garcia-s-hugging-face-page-presents-dale-8a0293"
title: "Manuel García’s Hugging Face page presents dale-0.8b"
description: "Manuel García’s Hugging Face page presents dale-0.8b, a 529 MB quantized model that turns English or Spanish requests into a single Linux Bash command."
retrieval_nugget: "Manuel García’s Hugging Face page presents dale-0.8b, a 529 MB quantized model that turns English or Spanish requests into a single Linux Bash command. Its author reports 79.4% accuracy on a 471-case test, ahead of larger untuned baselines; the practical appeal is local CPU use, while the benchmark remains author-run and task-specific."
published_at: "2026-11-09"
updated_at: "2026-11-09"
record_date: "2026-11-09"
date_kind: "published_at"
topics: ["local-models","models","open-models"]
entities: ["Manuel García","dale-0.8b-GGUF","dale","Qwen3.5-0.8B","Qwen3.5-9B","Gemma 4 E4B-it","llama.cpp"]
related_patterns: ["agent-security-moves-to-runtime-boundaries"]
novelty_label: "new_fact"
what_changed: "A new 800-million-parameter model for turning text into Linux Bash commands is reported to outperform Qwen 9B and Gemma 4."
prior_refs: ["https://newruntime.com/posts/a-community-publisher-has-released-an-altered-deepseek-9add4b/"]
source_url: "https://huggingface.co/manuelgscom/dale-0.8b-GGUF"
source_urls: ["https://huggingface.co/manuelgscom/dale-0.8b-GGUF"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:manuel-garcia-s-hugging-face-page-presents-dale-8a0293"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/manuel-garcia-s-hugging-face-page-presents-dale-8a0293/","markdown":"https://newruntime.com/posts/manuel-garcia-s-hugging-face-page-presents-dale-8a0293.md","json":"https://newruntime.com/posts/manuel-garcia-s-hugging-face-page-presents-dale-8a0293.json"}
markdown_url: "https://newruntime.com/posts/manuel-garcia-s-hugging-face-page-presents-dale-8a0293.md"
---

# Manuel García’s Hugging Face page presents dale-0.8b

## Retrieval answer

Manuel García’s Hugging Face page presents dale-0.8b, a 529 MB quantized model that turns English or Spanish requests into a single Linux Bash command. Its author reports 79.4% accuracy on a 471-case test, ahead of larger untuned baselines; the practical appeal is local CPU use, while the benchmark remains author-run and task-specific.

## Summary
Manuel García’s Hugging Face page presents dale-0.8b, a 529 MB quantized model that turns English or Spanish requests into a single Linux Bash command. Its author reports 79.4% accuracy on a 471-case test, ahead of larger untuned baselines; the practical appeal is local CPU use, while the benchmark remains author-run and task-specific.

## What happened
The Hugging Face repository describes dale-0.8b as a full fine-tune of Qwen3.5-0.8B. It is designed to convert a natural-language request into one Bash command for Linux and return a small JSON object with a type, command, and explanation. The possible types are `command`, `refuse`, and `unsure`: the model can decline a request outside shell tasks or signal ambiguity rather than inventing a command. The author says it was trained for this format and is not a general chat model.

The published artifact is a GGUF Q4_K_M quantization, with imatrix, sized at 529,297,952 bytes; the page lists Apache-2.0 as the license. The full-precision BF16 safetensors checkpoint is listed separately. The model can be run locally with llama.cpp, including through an OpenAI-compatible local server. The page also includes instructions for Ollama, vLLM, Docker Model Runner, and other clients. The author describes dale, the companion CLI, as an offline tool that uses the model on CPU after the first download. A suitable GPU can be used with larger family members; since dale 1.1.0, the CLI can select dale-4b or dale-9b when appropriate, while dale-0.8b remains its CPU model and fallback.

The author reports results on dale-bash-test-v2, a 471-case set: 339 command requests, 101 out-of-scope requests, and 31 ambiguous requests, in Spanish and English. dale-0.8b scored 374 correct, or 79.4%. On the same reported comparison, untuned Qwen3.5-9B scored 62.8%, as did Gemma 4 E4B-it; untuned Qwen3.5-0.8B scored 14.9%. The page gives a 95% Wilson interval of 75.5–82.8% for dale-0.8b and says the models were evaluated in Q4_K_M with reasoning disabled and greedy decoding. The author reports a 1.72-second median per request on a mid-range 2020 laptop CPU with four cores and eight threads, without a GPU.

The repository distinguishes the model from dale’s safety controls. Its diagram and description say the model proposes a command, while a deterministic C++ verifier classifies command text as safe, caution, or danger. The CLI offers suggest, confirm, and auto modes; in auto mode only commands classified safe run directly, while dangerous commands require typing a confirmation word. The author explicitly cautions that a standalone deployment outside dale does not include that verifier and says to review commands before running them.

## What the archive adds
New Runtime’s earlier coverage of open model checkpoints includes [Baidu’s publication of ERNIE-5.5-Flash-Base](https://newruntime.com/posts/baidu-has-published-open-weights-for-ernie-5-20af28/) (2026-11-05) and a [community-modified Qwen3.5-9B release](https://newruntime.com/posts/a-community-publisher-has-released-an-altered-deepseek-9add4b/) (2026-11-08). Those are related by distribution of model weights, not by a shared capability or benchmark. dale-0.8b is a new event: a specialized command-generation model with a local CPU-oriented use case.

## Trend signal
This release weakly touches [agent security moving to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/): dale separates command generation from a deterministic verifier and gates execution by risk. That is a concrete runtime boundary, but the source describes one CLI design, not evidence that this pattern is becoming broadly adopted. The GGUF alone lacks the verifier, so its safety properties should not be attributed to the model.

## Why now
The page situates dale-0.8b within a four-size family and says dale 1.1.0 can use larger siblings when a suitable GPU is present. That makes this release useful as the family’s CPU model and fallback, rather than a claim that one checkpoint fits every machine.

## Why it matters
For teams that need a local natural-language interface to routine Linux commands, a small model may be easier to run near the machine than a larger general-purpose model or a hosted API. The reported footprint and latency are specific enough to motivate a local trial, and the explicit `refuse` and `unsure` outputs address two common failure modes: answering outside the tool’s scope and guessing when a request is unclear.

The result does not establish that small fine-tuned models are generally better at shell work. It says this model did well on one published test under its own output contract. Teams should compare it against their actual commands and operating-system context, including less common utilities and destructive operations. More importantly, generation and execution are separate engineering problems. The companion CLI’s verifier and confirmation modes are part of the safety story; using only the weights does not reproduce those controls. Treat the model as a command suggestion component, not as authorization to execute shell text.

## Limits and uncertainty
The reported accuracy is from the model author’s test and evaluation setup; no independent replication is described on the repository page. The comparison also uses different prompt contracts for fine-tuned dale and untuned base models, which the author explains, so the figures measure complete configurations rather than isolating model size or fine-tuning alone. The test covers 471 cases, not the full range of Linux distributions, shell environments, permissions, or operational hazards. The reported latency is one CPU configuration. A deterministic verifier can classify command text, but the page’s description does not establish that it catches every harmful command or protects systems outside its CLI workflow.

## Reader takeaway
dale-0.8b is a practical, small local model for turning English or Spanish requests into Linux commands, and its author reports a strong result on a focused test. The important implementation detail is that safe execution belongs to dale’s separate verifier and confirmation flow, not to the GGUF by itself. Evaluate it on your own tasks, and inspect every generated command before execution.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Local Models](https://newruntime.com/topics/local-models.md): Explore the local-models topic hub.
- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)