LM Studio: complete guide 2026

#Why LM Studio in French

LM Studio checks the boxes that matter to get started: no command line, automatic GPU detection, model management from the interface, and a local OpenAI-compatible server usable from any app. For a French-speaking audience, two questions come up repeatedly: is the interface available in French, and where can you find models that respond properly in French.

The quick answer: yes for the interface (French is among the official languages), and yes for the models—Mistral Small, Qwen 3.5 / 3.8, and Gemma 4 handle French very well. The remaining question is which ones to load based on your VRAM and what you want to do.

#1. Set the interface to French

LM Studio starts by default in the system language. If your OS is in French, this has probably already been done. Otherwise, the change takes ten seconds.

1. 01
   Open settings

   Click the gear icon in the lower-left corner of the main window, or use the shortcut Ctrl+, (Cmd+, on macOS).
2. 02
   General section → Language

   In the Settings panel, find the General section. The Language menu lists the available languages.
3. 03
   Select French

   Select Français from the dropdown list. The interface switches immediately, without a restart.

#2. Interface tour

Once opened, LM Studio exposes five main areas through the left sidebar. Here's what each one is for.

Chat (bubble icon)

The main screen: conversation with the loaded model. Model selector at the top, settings on the right, and chat history on the left.

Discover / Search (magnifying glass icon)

Search for models on Hugging Face. Filter by architecture, size, and format (GGUF, MLX). This is where you download models.

My models (folder icon)

A list of everything downloaded to your disk, including size, quantization, and local path. Lets you uninstall items to reclaim space.

Local server (terminal icon)

Enables an OpenAI-compatible endpoint on http://localhost:1234. Essential for connecting third-party apps (VS Code, n8n, Python scripts) to LM Studio.

Settings (gear icon)

Language, theme, model folder, GPU backend selection (CUDA, ROCm, Metal, Vulkan), developer options.

#3. Find good French-language models

The Discover tab queries Hugging Face directly. For French-language use, a raw keyword search produces uneven results: many models claim to support French but hallucinate as soon as you move beyond simple topics. The method that works:

1. 01
   Filter by trusted publisher

   Type mistralai, Qwen, or google into the search field. These three families (Mistral, Qwen 3.5/3.8, Gemma 4) have native or near-native French. Avoid anonymous reuploads—prefer official accounts or bartowski / lmstudio-community for prepackaged GGUF versions.
2. 02
   Check the format

   LM Studio reads the GGUF format (and MLX on Apple Silicon). If the model has only raw safetensors weights, it cannot run directly. The model card indicates this.
3. 03
   Choose the right quantization

   Several files are often offered: Q4_K_M, Q5_K_M, Q8_0. Q4_K_M is the default balance. The quantization section below provides details.
4. 04
   Start the download

   Click the Download button to the right of the file. The bar at the bottom of the app shows the progress. You can continue using another model in the meantime.

#4. Recommended French-language models

Three families stand out for use in French in 2026. The right choice depends mainly on your available VRAM.

#Mistral (native French)

Edited by Mistral AI, a French company. French is treated as a first-class language, not an add-on. In 2026, the local offering focuses on the 24B tier, under the Apache 2.0 license:

Mistral Small 24B (Q4_K_M ≈ 14 GB)

The go-to French-language generalist for 16 GB+ of VRAM. Apache 2.0 license. Quality close to a proprietary model on complex French (writing, summarization, reasoning).

Devstral 24B (Q4_K_M ≈ 14 GB)

Same Mistral base, specialized for coding agents (Apache 2.0). If you code in French, it's the ideal companion — 16 GB+ of VRAM.

For less than 16 GB of VRAM, turn to Qwen or Gemma below: Mistral no longer has an up-to-date small general-purpose model in this range.

#Qwen 3.5 / 3.8 (strong multilingual performance)

Edited by Alibaba, under the Apache 2.0 license. Officially trained on dozens of languages, including French, with an unusually high level of consistency. This is the family that best covers all VRAM tiers, from CPUs to high-end cards.

Qwen 3.5 4B Instruct (Q4_K_M ≈ 3.4 GB)

The new “small default model.” Runs on 4–6 GB of VRAM or on the CPU. Surprisingly good in French for its size, ideal for classification/extraction.

Qwen 3.5 9B Instruct (Q4_K_M ≈ 6.6 GB)

THE 2026 8 GB choice: 256k of native context and vision (images). The sweet spot for 8–12 GB of VRAM (RTX 3060/4070). In Q8_0 (≈ 11 GB) on 12 GB, quality improves further.

Qwen 3.8 27B Instruct (Q4_K_M ≈ 18 GB)

For 24 GB of VRAM (RTX 3090/4090). Released on 14/08/2026: 262k context, vision, the closest to Copilot in the group. Tip: if it overthinks, set its reasoning to low in the settings.

#Gemma 4 (Google)

Since Gemma 4 (April 2026), the family has been available under the Apache 2.0 license and natively supports French, with multimodal capabilities (text + image) as a bonus. Often stylistically “cleaner” than older Llama models. Three sizes cover every budget: Gemma 4 E2B (Q4 ≈ 4.3 GB) for a small card or CPU use, Gemma 4 12B (Q4 ≈ 7.6 GB) multimodal for 8–12 GB of VRAM, and Gemma 4 26B-A4B (Q4 ≈ 19 GB), a fast multimodal MoE for 24 GB.

#5. Quantization explained

All GGUF models displayed in LM Studio are quantized. Understanding what that means keeps you from downloading 30 GB for nothing.

An LLM stores its weights as floating-point values. In the original version (FP16), each weight takes 2 bytes — so a 7B model takes about 14 GB. Quantization reduces the precision of these weights to save space, at the cost of a slight loss in quality.

Q4_K_M (recommended by default)

4 bits per weight on average. Cuts the size by ~4 vs. FP16. Minimal quality loss (~1-2% on benchmarks). This is the sensible choice in 90% of cases.

Q5_K_M

5 bits per weight. ~20% larger than Q4_K_M, with quality nearly indistinguishable from FP16. Choose it if you have spare VRAM and a demanding workload (code, long-form writing).

Q8_0

8 bits per weight. Nearly equivalent to FP16 in quality but 2× larger than Q4_K_M. Relevant for fine-tuning or reference benchmarks.

Q3_K_S / Q2_K

Extremely aggressive quantization. Noticeable quality loss on complex tasks. Reserve for cases where you absolutely need to fit a large model into limited VRAM.

FP16 (not quantized)

Native precision. 2 bytes per weight. Avoid it unless you have the VRAM and maximum quality is critical.

#6. Context and temperature

In the chat's right-hand panel, two settings have a far greater impact on the perceived experience than the others.

#Context Length (context length)

How many tokens the model can “see” at once—your prompt, the chat history, and the response. LM Studio defaults to 2048 or 4096, which is short: a moderately long conversation or a pasted document quickly exceeds the limit, and the model forgets the beginning.

Short conversations (general assistant)

4096 tokens are enough. Saves VRAM with no slowdown.

Document analysis (summarization, Q&A)

16k to 32k. Verify that the model natively supports this context (Qwen 3.5 9B: 256k, Qwen 3.8 27B: 262k, Granite 4.2 8B: 128k).

Code and long reports

32k+. Consumed VRAM rises quickly—typically +3 to +4 GB between 4k and 32k on a 7B.

#Temperature

Controls response randomness. The higher it is, the more risks the model takes when choosing the next word; the lower it is, the more it sticks to the most likely option.

0.0–0.3 (deterministic)

For data extraction, classification, code, or any task where you want the same answer every time. The model may seem dry or repetitive.

0.4–0.7 (balanced, default)

For general assistance, Q&A, and summarization. A good balance of consistency and variety. LM Studio often starts at 0.7.

0.8 – 1.2 (creative)

For freeform writing, brainstorming, and fiction. The model takes more stylistic liberties. Above 1.2, it quickly becomes incoherent.

#Tips and pitfalls

The model responds in English despite the French

Add an explicit system prompt: “You answer only in French, even if the question is asked in English.” Non-Mistral models sometimes need this.

Interrupted downloads

LM Studio does not always resume cleanly. If a download stalls, delete the partial file in My Models and restart it. Use a stable connection for large models (> 10 GB).

Unload a model to free VRAM

The Eject button at the top of the chat unloads the model. Useful before launching another GPU-intensive application (game, video rendering).

Multiple chats in parallel

Each chat keeps its own history and parameters. Useful for comparing two models side by side on the same question.

System prompt per model

The system prompt configured in the right-hand panel is saved with the conversation, not with the model. For persistent behavior, create a preset (Save Preset button).

MLX models on Mac

On Apple Silicon, LM Studio offers MLX versions in addition to GGUF. At the same size, MLX is faster. Mistral and Qwen have official MLX versions.

#Go further

You have LM Studio in French, you know how to choose a French model and tune the parameters that matter. The natural next directions are:

Enable the local server

The guide to turning LM Studio into an API server explains how to expose the OpenAI-compatible endpoint on http://localhost:1234 and connect it to VS Code, n8n, or a Python script.

Compare with Ollama

The Ollama vs LM Studio vs Jan vs GPT4All guide compares GUI/CLI tools. Many people combine Ollama (daemon) + LM Studio (client interface) to get the best of both.

Explore quantization in depth

The Choosing Your Quantization guide (Q4, Q5, Q8, FP16) visually compares the actual quality loss across tasks.