askmdb-9b-GGUF

askmdb-9b-GGUF

Writes a read-only MongoDB aggregate pipeline from a plain Spanish or English question. A GGUF copy of a Qwen3.5-9B fine-tune that runs locally with llama.cpp, for developers who want to query data in everyday language.

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: askmdb-9b is a 9 billion parameter model that writes MongoDB queries. You ask in Spanish or English, and it replies with a small JSON object holding a single read-only aggregate pipeline. It is Qwen3.5-9B with a LoRA fine-tune merged into its weights, packed here as a Q4_K_M GGUF file for llama.cpp. You paste a short summary of your collections and fields into the prompt, so it can handle a database it has not met before.

Who it suits: Developers, analysts and support staff who want to question a MongoDB database in everyday language without writing the pipeline by hand, and who prefer a model that runs on their own machine. It is the largest of four askmdb sizes and the one aimed at a graphics card.

Industry | Who uses it | How they use it | Example
Software | Backend developers | Draft aggregate pipelines from a question | Asking for the newest orders per customer
Data | Analysts | Query a MongoDB collection in plain Spanish or English | Counting sign-ups by month
Support | Support engineers | Look up records without learning pipeline syntax | Finding open tickets for one account
Education | Database students | Compare a question with the pipeline it produces | Reading how a group stage is built

You can skip it if you do not use MongoDB. If your graphics card has less than 8 GiB of free memory, the smaller askmdb sizes suit you better, since the command line tool offers this one only above that line.

Where to use it:

• The askmdb command line tool, which runs it on a graphics card through llama.cpp and Vulkan
• llama-server from llama.cpp, inside any other application, on CPU or GPU
• Windows, macOS or Linux with Python, to download the file

Use case | Who | What they get back
Plain-language questions | Developers | A JSON answer with one read-only pipeline and a short explanation
Checked querying in askmdb | Analysts | Only read pipelines run, and a join asks you first in auto mode
Your own application | Integrators | Pipelines from a local llama-server endpoint

How to set it up:

• Python, to install the Hugging Face client
• llama.cpp or the askmdb tool, to run the file
• About 5.6 GB of disk for the Q4_K_M file
• A graphics card with at least 8 GiB of free memory, if you want askmdb's GPU mode

1. Install the Hugging Face client.
2. Download the repository files to your computer.
3. Load the GGUF file in llama.cpp, or let the askmdb tool offer it when your GPU has enough free memory.
4. Connect to MongoDB with a read-only user, since nothing outside askmdb checks the query.

pip install -U huggingface_hub
hf download manuelgscom/askmdb-9b-GGUF

How to use it: Give the model a compact description of your collections and fields along with your question. It answers with a JSON object that holds the pipeline. Inside askmdb, a checker lets only read pipelines run. The source shows no serve command, so follow the llama.cpp guide for llama-server. Outside askmdb nothing verifies the pipeline, so read it before you run it.

Tool | What it does
askmdb | Command line tool that shows the pipeline, explains it and runs it after a safety check
llama-server | Serves the model to other applications, on CPU or GPU
Q4_K_M GGUF file | The compressed weights, 5,629.1 MB
Schema summary | A short description of your database that you put in the prompt

Example: You ask, in English or Spanish, how many orders each country has in a collection you described in the prompt. The model returns a JSON object with one aggregate pipeline and a short explanation. You read it, then run it with a read-only MongoDB user.

For newcomers and experts: If you are new, start with the askmdb tool, since it shows the pipeline, explains it and checks that it only reads before anything runs. Try one simple question first.

If you know the field: the weights are Qwen3.5-9B with the LoRA merged in, quantized to Q4_K_M with an importance matrix. All four askmdb sizes share the same data, output contract and prompt. Safety is not in the model. In askmdb a deterministic verifier does that job.

Question plus schema summary
askmdb-9b writes a pipeline
Verifier allows reads only
Query runs on MongoDB

Good to know: The authors scored the model on a public set of 500 cases by running each query. The result beats the base models they compared without fine-tuning.

Title: Test accuracy on 500 public cases, in percent
askmdb-9b: 91.6%
askmdb-0.8b: 83.6%
Qwen3.5-9B: 47.8%
Gemma 4 E4B-it: 45.4%

• Speed: about 1.85 s per question on the GPU, using 5.5 GiB of VRAM, inside the askmdb tool
• Languages: Spanish and English
• Safety: the model does not judge safety, so outside askmdb check each query and use a read-only user
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: model files downloaded to your computer, and the MongoDB database you point askmdb at when it runs the queries
• Not for: databases other than MongoDB, or writing and changing data, since it produces read-only pipelines

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 9B
• Weights: about 4.5 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 351
• Likes: 5
• License: apache-2.0
• Languages: es, en
• Based on: Qwen/Qwen3.5-9B
• Tags: gguf, mongodb, text-to-query, nl2query, llama.cpp, vulkan, text-generation, model-index, imatrix, conversational

Links

• Model: https://huggingface.co/manuelgscom/askmdb-9b-GGUF
• Files: https://huggingface.co/manuelgscom/askmdb-9b-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download manuelgscom/askmdb-9b-GGUF

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