Subscribe
0%
Illustration for: OVHcloud's Olala mini weights are out: first benchmarks, Apache 2.0, and why the '8 GB GPU' claim needs a footnote

OVHcloud's Olala mini weights are out: first benchmarks, Apache 2.0, and why the '8 GB GPU' claim needs a footnote

OVHcloud published the weights for Olala mini 7A1B, its first in-house language model, on Hugging Face on November 19. The card adds the quality scores that were missing at the November 17 announcement. They're solid but not class-leading, and running it locally currently needs OVHai's own forks of vLLM, llama.cpp and MLX.

Listen to this article0:00 / 5:08

OVHcloud has published the weights for Olala mini 7A1B, its first in-house large language model, on Hugging Face. The model card gives a November 19 release date, which is the day of the company's AI Summit in Paris. The license is Apache 2.0. Alongside the post-trained model, OVHai LLM (the OVHcloud unit built from its March acquisition of Dragon LLM, per ActuIA) released a base model, an SFT checkpoint and a technical report.

Why it matters: when OVHcloud announced the model on November 17, it gave speed and memory figures but no quality scores. ActuIA's headline summed it up as "speeds, no scores". The card now fills that gap. It's the first chance to see where a European-trained, Apache-licensed small model actually lands.

What it is

Per the model card:

Item Olala mini 7A1B
Parameters 6.8B total, about 1B active per token
Design 36 layers (29 Mamba-3, 7 attention), each with a mixture-of-experts block: 256 routed experts (6 active) plus 1 shared
Context 32K tokens, input plus output. A 128K version is "coming soon"
Training 12.8T tokens of pre-training on JUPITER, Jean Zay and MeluXina; base model trained on 1,024 GH200 GPUs
Data cutoff Pre-training January 2026, post-training April 2026
Reasoning Four effort levels, from disabled to high
Modalities Text only; English plus 36 other languages

OVHcloud says European public support for the training was worth about €7 million in computing hours. Search engine Qwant already uses the model for its quick-answer summaries. Qwant reports answers 20% faster at about a third of the previous running cost. Both are company claims. ActuIA notes that Qwant is 75% owned by the Klaba brothers, who founded OVHcloud.

The benchmarks

OVHai LLM ran every model itself, using the same harness and cluster. Olala ran at its highest reasoning effort. Its average across the suite is 68.0. A selection from the card's table (%):

Benchmark Olala mini Granite 4.0-H-Tiny LFM2.5 8B-A1B Ling 3.0-tiny
IFBench (instruction following) 57.7 14.3 48.3 57.3
Multi-IF 71.2 44.5 71.0 75.2
GPQA Diamond 50.5 30.3 41.9 68.7
HumanEval+ 88.3 77.4 79.9 92.0
AIME 2025 61.3 4.7 48.0 70.0
BFCL V3 (tool calling) 59.6 60.2 69.7 68.3

These numbers are OVHcloud's own, not independent results. Read that way, Olala clearly beats IBM's Granite tiny model on most rows and roughly trades blows with Liquid's LFM2.5 8B-A1B. On almost every row shown, though, it trails Ling 3.0-tiny. The card itself calls competition maths its weakest area and instruction following its strongest. On safety testing, 6.3% of attack prompts got through, and about one in ten harmless requests was refused.

The "8 GB GPU" claim, in practice

OVHcloud's announcement said the model runs on a GPU with as little as 8 GB. It also claimed 160 tokens/s on an RTX A4500 and 50 tokens/s on a standard CPU. ActuIA pointed out that the A4500 has 20 GB and that BF16 weights alone come to roughly 14 GB. The cards OVHai published today help explain the gap:

  • vLLM needs OVHai's fork or Docker image and an NVIDIA GPU with 24 GB or more (tested on L4 and L40S). Official vLLM can't load the model yet.
  • Quantized GGUF files range from 6 to 14 GB and are meant for a Mac M1, a PC without a GPU, or a single NVIDIA card. That's where the small-memory claim plausibly comes from. They also need OVHai's llama.cpp fork.
  • Apple Silicon: the MLX card lists 4.9 GB mixed-precision builds and reports about 64–86 tokens/s on a 16 GB MacBook Air M5 in OVHai's own tests, depending on the build. It requires OVHai's mlx-engine fork.
  • Plain Transformers on CPU is slow: the card cites about 6 tokens/s on a 64-core server without extra kernels.
  • Ollama and LM Studio can't load it yet, according to the card.

The card also warns that generic quantization recipes can produce NaN output, so stick to the published files. Ask This Guy, citing French outlet Next, notes that training code and data aren't released. "Open-weight" is the accurate label, not fully open source.

What to watch

  • Independent evals. Until the weights were out, nobody outside OVHcloud could test the model. Third-party runs, especially against Qwen and Ling models of similar size, will show whether the card's numbers hold.
  • Upstream support. Day-to-day usability depends on mainline vLLM, llama.cpp and MLX adding the Mamba-3/MoE architecture.
  • Version 1.1, promised by the end of 2026 with 128K context and better maths. A version 2.0 above 100B parameters has no date.

Who should try it now: teams that want an Apache-licensed small model trained in Europe for summarisation, instruction following or multilingual European text, and are comfortable running a forked inference stack. Anyone who needs a drop-in Ollama model should wait.

Sources

  1. Hugging Face — OVHaiLLM/Olala-v1.0-mini-7A1B model card huggingface.co
  2. Hugging Face — OVHaiLLM/Olala-v1.0-mini-7A1B-MLX model card huggingface.co
  3. OVHcloud — OVH announces Olala mini 7A1B corporate.ovhcloud.com
  4. ActuIA — OVHcloud unveils Olala mini, its first LLM: speeds, no scores actuia.com
  5. Ask This Guy — Olala mini, OVH's first French LLM: what is it really worth? askthisguy.com
  6. Compare the Cloud — OVHcloud releases Olala mini, an open-source language model comparethecloud.net

#OVHcloud#Olala#Olala mini 7A1B#open-weight models#mixture of experts#Mamba#Hugging Face#Qwant#Europe#sovereign AI