Skip to content

THE FIELD JOURNAL

Olala mini: Can You Run OVHcloud's Open Model Locally?

By Local AI Scout · Beginner guide

Paper-cut illustration of an open laptop with a blank screen on a wooden desk, beside a small paper lighthouse and a stack of folded maps under a desk lamp, with paper hills at dusk through the window

By Local AI Scout · Local · Models · Guide · Docs-only · Checked November 19, 2026

Paper-cut illustration of an open laptop with a blank screen on a wooden desk, beside a small paper lighthouse and a stack of folded maps under a desk lamp, with paper hills at dusk through the window

Short answer: Olala mini (full name Olala-v1.0-mini-7A1B) is an open-weight model from OVHai LLM, part of the French cloud company OVHcloud. It has 6.8 billion parameters, about 1 billion of them active per token, a 32K-token context and an Apache 2.0 license. OVHcloud publishes its own GGUF files from 6.0 GB and MLX folders from 4.9 GB, so it's sized for a 16GB laptop. The catch: OVHcloud says Ollama, LM Studio and the official llama.cpp can't load it yet. Today you run it with OVHcloud's own fork of llama.cpp or of LM Studio's MLX engine, or through a hosted API.

This guide follows one reader: a freelance consultant with a 16GB MacBook Air who already runs models in Ollama and wants to know whether Olala is worth trying for client notes that shouldn't leave her laptop.

What Olala mini is

OVHcloud's model card describes it as "a small reasoning model for chat, instruction following and tool use". The parts that matter for running it at home:

  • Size: 6.8B parameters in total, about 1B active per token. It's a mixture-of-experts model, so it does less work per token than its size suggests, but you still download all of it.
  • Architecture: 36 layers, 29 of them Mamba-3 and 7 attention. OVHcloud says only those 7 layers keep a KV cache, which is part of why it fits on a laptop.
  • Context: 32K tokens, input and output together. OVHcloud says a 128K version is "coming soon".
  • Reasoning: it thinks before it answers, with four effort levels (disabled, low, medium, high), and it was trained to call tools.
  • Made in Europe: designed in France and trained on three European supercomputers (JUPITER, Jean Zay and MeluXina) on 12.8 trillion tokens.
  • Knowledge cut-off: January 2026 for pre-training (April 2026 for post-training). It has no clock, so give it the date when it matters.
  • License: Apache 2.0. The license table lists the release date as November 19, 2026, and says to keep the NOTICE file if you redistribute it.

Why Ollama and LM Studio can't load it yet

Olala uses its own architecture, called olala, and support hasn't reached the usual apps. OVHcloud's model card is direct about it: "Ollama, LM Studio and the official llama.cpp, vLLM and mlx-lm cannot load Olala 1.0 mini yet". The GGUF page adds that the apps under Hugging Face's "Use this model" button can't load the files either.

So our consultant can download the files, but ollama run won't work, and there was no olala model in Ollama's library when we checked. OVHcloud hasn't said when support will reach Ollama or LM Studio ‡.

Which file fits your computer

These are OVHcloud's own files and descriptions:

FormatFile or folderSizeOVHcloud says use it for
GGUFBF1613.6 GBNVIDIA GPU; Mac with 24 GB or more
GGUFq8_0-safe2 (8-bit)6.9 GBx86 CPU; Mac with 16 GB
GGUFq6_k (6-bit)6.3 GBx86 CPU; Mac with 16 GB (fastest quantized file on Metal)
GGUFq5_k_m (5-bit)6.0 GBx86 CPU, Mac; smallest file with no measured quality loss
MLXq87.2 GBApple Silicon
MLXq5e6 or q4e84.9 GBApple Silicon; OVHcloud says q5e6 uses 5.4 GB on a 16 GB MacBook Air M5

Two things are different from most models:

  • There's no 4-bit GGUF. The usual advice to grab Q4_K_M doesn't apply. OVHcloud says it built and tested a Q4_K_M file, then withdrew it because it saved only 0.3 GB over q5_k_m and was "measurably worse".
  • Don't make your own quants. OVHcloud warns that generic quantization recipes "produce NaN on realistic inputs" on this model. Use the published files.

OVHcloud reports about 72 tokens per second for q6_k and about 86 for the MLX q5e6 folder on a 16 GB MacBook Air M5. Those are OVHcloud's measurements; we didn't run the model.

How to run it today

There are three ways in, all from OVHcloud's model cards:

  1. On a Mac with MLX (fastest on Apple Silicon, according to OVHcloud). Clone OVHcloud's fork of LM Studio's mlx-engine, install it with Python 3.11 on macOS 14 or later, download one folder (for example q5e6) and chat from Python or start its OpenAI-compatible server. This doesn't put Olala inside the LM Studio app.
  2. With llama.cpp (Mac M1, Windows or Linux PCs, NVIDIA GPUs). Build OVHcloud's fork of llama.cpp from GitHub, download one GGUF file, then use llama-cli to chat or llama-server for an API. The build step is the same as in our llama.cpp install guide, just with OVHcloud's repository.
  3. Hosted, with nothing to install. OVHcloud serves it through its AI Endpoints service, and the card also lists Blade. That isn't local: your prompts go to their servers.

For Apple M1 owners, both cards give an extra setting to export before you start: OLALA_NO_FUSED_MOE=1 for llama.cpp (OVHcloud says it makes generation about 60% faster) and OLALA_MOE_V3=0 for MLX (without it, M1 Macs crash on the first token).

What it's good at, and where it's weak

OVHcloud names its intended uses as chat in "English and the main European languages", tool-calling agents, reasoning in mathematics, science and code, and "local and on-premises deployment, when data must stay on your own machines". It's equally plain about the limits:

  • Facts from memory. With about 1B active parameters "it stores less factual knowledge than larger models and can state wrong facts with confidence". Give it the document or a search tool.
  • Other languages. Its comprehension outside English and the main European languages "is weaker than its training mix suggests".
  • Math. Competition mathematics is its weakest benchmark area.
  • Safety. In OVHcloud's tests 6.3% of attack prompts still got through, and it refuses about 1 benign request in 10.

Try Olala mini now if…

  • You're comfortable cloning a GitHub repository and running commands in Terminal.
  • You want an Apache 2.0 model made in Europe that fits a 16GB machine.
  • Your work involves tool calling or step-by-step reasoning, and you'll check its facts.

Wait for Ollama or LM Studio support if…

  • You only use apps with a model menu. That was our consultant, so for now she's sticking with a model she can pull in Ollama on her Mac.
  • You need more than 32K tokens of context: OVHcloud says a 128K version is coming.
  • You'd rather pick from models that already run everywhere; see our model guide.

Frequently asked questions

Can I run Olala mini in Ollama?

Not yet. OVHcloud's model card says Ollama, LM Studio and the official llama.cpp can't load Olala 1.0 mini yet, and there was no olala model in Ollama's library on November 19, 2026. Use OVHcloud's llama.cpp or mlx-engine fork instead.

How much memory does Olala mini need?

OVHcloud's smallest GGUF file is 6.0 GB and its smallest MLX folders are 4.9 GB, and it says its 6-bit GGUF file and its q5e6 MLX folder (which uses 5.4 GB) run on a 16 GB Mac. The full BF16 version is 13.6 GB and needs a Mac with 24 GB or more, or an NVIDIA GPU.

Is there a 4-bit (Q4_K_M) version of Olala mini?

No. OVHcloud built and tested a Q4_K_M file, then withdrew it because it saved only 0.3 GB over the 5-bit file and was measurably worse. It also warns that generic quantization recipes produce NaN on this model, so use the published files.

Can I use Olala mini for commercial work?

It's released under the Apache 2.0 license, which allows commercial use, and the card asks you to keep its NOTICE file when you redistribute it. If you build a product on it, read what commercial use of local models involves.

Who makes Olala mini?

OVHai LLM, part of the OVHcloud group. The model card says it was designed in France and trained on three European supercomputers.

Sources

Checked November 19, 2026. Docs-only: we read OVHcloud's pages above and didn't download or run the model. Speeds and quality figures are OVHcloud's own. See our editorial policy.

‡ = not stated on an official page we opened, or our own reading rather than the source's words. Check before relying on it.

Keep exploring: Apps · Models · All guides

THE SCOUT LETTER

Good finds.
Less noise.

Useful local AI apps, practical guides and the updates worth your time. A thoughtful note in your inbox.

Free to join. Unsubscribe with one click.