Ollama Modelfile: create and customize your model

#Why use a Modelfile?

When you type ollama run qwen3.5:9b, you get a generic model. It sometimes answers a question in French in English, comments on its code answers with endless introductions, and negotiates when you ask for a literal translation. Instead of repeating the same system prompt in every session, you freeze it once in a Modelfile and get a reusable variant.

In practice, a Modelfile does three things: (1) attaches a persistent system prompt to the model, (2) configures generation parameters (temperature, context, stop tokens), and (3) optionally replaces the chat template. The created variant is used like any Ollama model: ollama run mon-assistant.

#Prerequisites

Ollama installed

Version 0.3+ recommended. Check with ollama --version. The daemon must be running on http://localhost:11434.

An already pulled base model

For example, ollama pull qwen3.5:9b or ollama pull qwen3-coder:30b. The Modelfile inherits from this model.

A text editor

VSCode, Notepad++, vim — any of them. The file does not require an extension; by convention, it is generally called Modelfile.

Approximately 5 minutes

It takes just long enough to create the file and launch ollama create. No additional download: an already-present model is reused.

#1. Modelfile syntax

A Modelfile is a text file with uppercase instructions, one per block. The order does not matter, but convention places FROM first.

FROM qwen3.5:9b

SYSTEM """
Tu es un assistant utile et concis.
"""

PARAMETER temperature 0.7
PARAMETER num_ctx 4096

The main instructions:

FROM

Base model. Required. Accepts a Ollama name (qwen3.5:9b, qwen3.8:27b) or a local path to a GGUF.

SYSTEM

The system prompt injected into every conversation. Use triple quotes for multiline text.

PARAMETER

Sampling and context settings. One instruction per parameter. The most useful ones: temperature, top_p, top_k, num_ctx, repeat_penalty, num_predict, stop.

TEMPLATE

Full prompt format (rare). Only change it if you know what you're doing—each model family has its own template.

MESSAGE

Few-shot examples. Prefixed with MESSAGE user or MESSAGE assistant. Useful for anchoring a style.

ADAPTER

Path to a GGUF LoRA to apply on top of FROM. For advanced users.

LICENSE

Free text, for traceability. Does not affect execution.

#2. Create your first model

The workflow is always the same: write a Modelfile, run ollama create, and test with ollama run.

1. 01
   Create the file

   In a working directory, create a file named Modelfile (with no extension). Put your variant definition in it.
2. 02
   Start creation

   From the same folder: ollama create mon-modele -f ./Modelfile. Ollama validates the syntax, inherits the model weights from FROM, and registers a new entry. It is instantaneous and does not download anything again.
3. 03
   Check

   ollama list montre votre nouveau modèle aux côtés des autres. La taille affichée est identique à celle du modèle de base — c'est une référence, pas une copie.
4. 04
   Test

   ollama run mon-modele. Le system prompt et les paramètres sont déjà appliqués. Vous pouvez aussi pointer une interface (Open WebUI, LM Studio) sur cette variante via l'API.

# Dans le dossier qui contient le Modelfile
ollama create assistant-fr -f ./Modelfile
ollama list
ollama run assistant-fr

#3. Example: concise French assistant

Goal: an assistant that consistently responds in French, without preambles such as "Of course, here’s…", without apologetic formulas, and with calibrated length.

FROM qwen3.5:9b

SYSTEM """
Tu es un assistant francophone précis et concis.

Règles strictes :
- Réponds toujours en français, jamais en anglais, même si l'utilisateur écrit en anglais.
- Pas de préambule ("Bien sûr", "Voici", "Absolument"). Va directement au fait.
- Pas d'excuses ni de disclaimers sauf si l'information est réellement incertaine.
- Réponses courtes par défaut (3-5 phrases). Détaille uniquement si on te le demande explicitement.
- Si tu ne sais pas, dis-le en une phrase. N'invente pas.
"""

PARAMETER temperature 0.6
PARAMETER top_p 0.9
PARAMETER num_ctx 8192
PARAMETER repeat_penalty 1.15

Creation and testing:

ollama create assistant-fr -f ./Modelfile-assistant-fr
ollama run assistant-fr "Quels sont les avantages d'un LLM local ?"

A repeat_penalty of 1.15 (instead of the default 1.1) stops models from repeating themselves. A num_ctx of 8192 lets you paste somewhat longer documents without breaking the default 2048-token context.

#4. Example: quiet Python coding

Goal: a coding companion that returns Python ready to paste, with minimal surrounding prose. Ideal for workflows where you want to pipe the output into an editor or script.

FROM qwen3-coder:30b

SYSTEM """
Tu es un assistant de code Python expert. Tu suis ces règles :

1. Réponds uniquement avec du code Python valide, dans un bloc fenced ```python.
2. Pas d'explication avant ou après le code, sauf si l'utilisateur demande explicitement une explication.
3. Le code doit être complet, exécutable, avec les imports nécessaires en haut.
4. Privilégie la stdlib quand c'est possible. Si une dépendance externe est nécessaire, mets un commentaire en tête : # pip install <package>.
5. Type hints quand c'est raisonnable (Python 3.10+).
6. Si la demande est ambiguë, choisis l'interprétation la plus probable et code-la, sans poser de questions.
"""

PARAMETER temperature 0.2
PARAMETER top_p 0.95
PARAMETER num_ctx 16384
PARAMETER stop "```\n\n"

A few notable technical choices:

FROM qwen3-coder:30b

A code-specialized model from 2026 (MoE 30B-A3B, only 3B active parameters, 256k context, ~19 GB in Q4). On a 24 GB card (RTX 4090) or a 32 GB Mac, it runs at 50–80 tokens/s thanks to its 3B active parameters. If you only have 16 GB of VRAM, replace it with gpt-oss:20b (14 GB) or devstral:24b.

temperature 0.2

Low for coding. At 0.7 (default), the model improvises variable names and sometimes invents APIs. For code, you want determinism.

num_ctx 16384

Extended window for analyzing entire files. Warning: it multiplies the VRAM consumed by the context.

stop

The model stops after the first code block. This avoids post-code explanations cluttering the output.

ollama create coder-py -f ./Modelfile-coder-py
ollama run coder-py "Lire un CSV avec pandas et retourner la moyenne de la colonne 'prix'"

#5. Example: strict EN→FR translator

Goal: a translator that returns ONLY the translation, without quotation marks, without "Translation:", and without stylistic variation. The typical use case: integration into an n8n pipeline or a bash script where the output will be consumed raw.

FROM qwen3.5:9b

SYSTEM """
Tu es un traducteur professionnel anglais → français.

Règles ABSOLUES :
- Tu reçois un texte en anglais. Tu retournes UNIQUEMENT la traduction française.
- Pas de préfixe ("Traduction :", "Voici :"), pas de guillemets autour de ta réponse, pas de commentaire.
- Conserve la mise en forme exacte du texte source (sauts de ligne, listes, ponctuation).
- Conserve les noms propres, marques, URLs et identifiants techniques tels quels.
- Si le texte est déjà en français, renvoie-le inchangé.
- Si le texte est dans une autre langue que l'anglais, renvoie : ERREUR_LANGUE_NON_SUPPORTÉE
"""

PARAMETER temperature 0.3
PARAMETER num_ctx 4096
PARAMETER num_predict 2048

Adding MESSAGE to anchor the format through few-shot prompting makes the behavior even more predictable:

FROM qwen3.5:9b

SYSTEM """
Tu traduis de l'anglais vers le français. Tu renvoies UNIQUEMENT la traduction, sans préfixe ni guillemets.
"""

MESSAGE user "The meeting starts at 3 PM."
MESSAGE assistant "La réunion commence à 15 heures."

MESSAGE user "Please check the attached document."
MESSAGE assistant "Merci de consulter le document ci-joint."

PARAMETER temperature 0.3

Pipeline usage:

echo "The deployment is scheduled for tomorrow morning." | ollama run translator-en-fr

#6. Managing custom models

Over time, you'll accumulate several variants. A few useful commands to stay in control:

# Lister tous les modèles (originaux + variantes)
ollama list

# Voir le Modelfile d'une variante (utile pour récupérer un Modelfile perdu)
ollama show --modelfile assistant-fr

# Voir les paramètres effectifs
ollama show --parameters assistant-fr

# Voir le system prompt seul
ollama show --system assistant-fr

# Supprimer une variante (libère uniquement l'entrée, pas les poids du modèle de base)
ollama rm assistant-fr

#7. Which base model should you choose?

Every Modelfile starts with FROM: choosing the base model determines 90% of the final result. A perfect system prompt can never make up for a base that is poorly suited to your use case or too large for your VRAM.

General-purpose assistant

A recent 8B to 14B model (from the Qwen or Gemma family) in Q4: responsive, good in French, and leaves room for context.

Code

A code-specialized model—MoE models such as Qwen3-Coder 30B-A3B are very fast with 20 GB of memory or more; below that, a 7–9B code model remains a good choice.

Formal French / writing

Models Mistral (Nemo, Magistral) retain the advantage in French at the same size.

Small configuration (8 GB of VRAM)

Stick to 4B to 9B in Q4: a model that spills into RAM ruins responsiveness, regardless of the Modelfile.

To compare models by VRAM, license, and use case, the QuelLLM catalog filters models compatible with your machine—each entry provides the ollama run exacte command to put in your FROM.

#Troubleshooting

The model ignores the system prompt

Make sure you're actually launching your variant (ollama run mon-modele) and not the base model. Confirm with ollama show --system mon-modele.

Truncated responses

Increase num_predict (default 128 in some configs) to 2048 or higher. Or add PARAMETER num_predict -1 to disable the limit.

The model does not follow a strict rule

Restate the rule in uppercase, add "ABSOLUTELY" or "NEVER", and add an example via MESSAGE. Small models (under ~10B) struggle with isolated negations.

Error "Error: invalid model reference"

The FROM model is not installed locally. Run ollama pull <modele> before the ollama create.

VRAM saturated after creation

A high num_ctx multiplies consumption. If you go from 2048 to 16384, expect +1 to +3 GB of VRAM depending on the model size. Drop back to 8192 if memory gets tight.

#Go further

The Modelfile is the basic building block for specializing Ollama without touching the weights. Three natural directions to explore next:

Mastering system prompts in greater depth

The guide to system prompts explains how to structure roles, constraints, and examples for truly reliable behavior—useful as soon as you exceed 5 lines of SYSTEM.

Adjust temperature, top-p, and top-k

Before touching the model, knowing precisely what each sampling parameter changes avoids a lot of back-and-forth. The dedicated guide covers all useful settings.

Expose your variants through Open WebUI

Once you create your custom models, Open WebUI lists them automatically. You can switch between assistant-fr, coder-py, and translator-en-fr with one click from the interface.