Skip to content

Metis Hybrid Olala: a European decision model in one day

The Metis wordmark in white on black, with a green dot on the i, followed by Hybrid Olala in grey, above the line from GreenPT.

OVHcloud launched Olala mini on 19 November. A day later we had trained Metis Hybrid Olala on it, one of very few decision models on a European base.

You can run a calibrated decision model whose base model was designed and trained in Europe. OVHcloud launched Olala mini at its AI Summit on 19 November. On 20 November we trained Metis Hybrid Olala on top of it, and on 21 November the weights and the code to run them went public on Hugging Face.

What is Metis Hybrid Olala?

Metis Hybrid Olala is a decision model: you give it a situation and a question with fixed options, and it returns a calibrated probability for each option in one forward pass. It is built on OVHcloud’s Olala mini and speaks the same /v1/systemone API as TypeSafe’s Jev, so Jev clients work unchanged.

It joins Metis Hybrid, our family of decision models that also write. “Hybrid” means one model does two jobs. A small pointer head reads the model’s hidden state and picks an option, so a refund check or a ticket route comes back as numbers you can log and threshold. The same model also writes text, so it can explain a decision or answer a chat message.

One model, two jobs: the pointer head decides, the same model writes

A real request to Metis Hybrid Olala and its answer, 21 November 2026

REQUESTRefunds require a receipt.The customer has no receipt.What should the agent do?refunddeclineONE MODELOlala mini + Metis adapter~1B of 6.8B activePOINTER HEADdecline0.92refund0.08LANGUAGE MODELexplains the decision,or chatsModes: decision, both, text

Olala mini is a hybrid too, in a different sense. OVHcloud built it from Mamba-3 and attention layers with a mixture of experts: 6.8B parameters in total, of which about 1B are active per token. It was trained on 12.8T tokens on three European supercomputers: JUPITER in Germany, Jean Zay in France and MeluXina in Luxembourg.

Why so few decision models are European

Almost every open decision model sits on a base model from China or the United States. On the community Decision Index, which tests decision models on the same hardware, we counted the base models behind all 217 entries of about 10B parameters or fewer (data of 21 November 2026).

  • 138 build on Alibaba’s Qwen.
  • Most of the rest build on US models such as Google’s Gemma, Liquid AI’s models and the Ettin and ModernBERT encoders.
  • 3 use a European base model: StandardOne 3B and 8B by Standard Thinking, and Jevstral 8B, an independent project. All three build on Mistral’s Ministral 3.

3 of 217 small decision models use a European base model

Base models of Decision Index entries of about 10B parameters or fewer, 21 November 2026

138on Alibaba's Qwen76on other bases, mostly from the US3on a European base (Mistral'sMinistral 3)+1 Metis Hybrid Olala, on OVHcloud's Olalamini (not on the index yet)

If your organisation wants its AI supply chain in Europe, that is a thin shelf. Metis Hybrid Olala adds a fourth option, and to our knowledge it is the first decision model built on Olala.

One day from launch to decision model

Metis Hybrid Olala took one working day because the recipe already existed. It is the same round-2 recipe as our Metis Hybrid 4B, 9B and 27B, with the same data and settings. Only the parts that touch the backbone had to change.

From Olala's launch to a public decision model in two days

November 2026

19 NovOVHcloud launches Olala mini at its AISummit20 Nov, one day laterWe train Metis Hybrid Olala: about 4.5hours on one H20021 NovWeights and code public on Hugging Face

Here is how 20 November went:

  1. Olala mini wrote the chat answers for our training mix itself, through OVHcloud’s vLLM build, so the model keeps its own voice in chat.
  2. We pointed the adapter at Olala’s own layers: the linear layers of its Mamba-3 and attention mixers and the projections around its experts. Olala’s chat format uses channels, so decision prompts open the answer channel directly and get a short system prompt instead of Olala’s default of about 700 tokens.
  3. Stage 1 trained on 136,269 rows for two epochs, in 224 minutes on one H200 GPU.
  4. Stage 2 trained on 42,391 rows of code-built hard tasks plus a replay of stage 1, for one epoch, in 45 minutes.
  5. We fitted the per-input calibrator on our own evaluation sets, never on JevBench.

That is about 4.5 hours of GPU training in total. The model was in our Hugging Face organisation by the end of the day.

How it scores

On the 231 public JevBench questions (version 1.2), Metis Hybrid Olala answers 169 correctly, the same score its model card reports. That puts it level with StandardOne 3B and Jevstral 8B, even though it uses a third or less of their parameters per token. We trained on none of these questions.

We ran every European-base decision model through the same questions, each as its model card says to serve it:

Model Base model Parameters per token Right of 231 Hard (111)
Metis Hybrid Olala Olala mini (OVHcloud) about 1B 169 62
StandardOne 8B Ministral 3 (Mistral) 8.4B 182 63
StandardOne 3B Ministral 3 (Mistral) 3.4B 165 51
Jevstral 8B Ministral 3 (Mistral) 8.4B 161 56

StandardOne’s own cards report 182 and 166, so our runs match what their builders published. Jevstral cannot answer the 18 rating-scale questions, which count as wrong.

Level with StandardOne 3B and Jevstral 8B at about 1B parameters per token

JevBench public, 231 questions, our runs. The vertical axis starts at 155.

Right of 23118517516515502468Parameters used per token (billions)StandardOne 3B165 · 3.4BJevstral 8B161 · 8.4BStandardOne 8B182 · 8.4BMetis Hybrid Olala169 · about 1B

Three things stand out:

  • Level with models three to eight times its active size. Our paired test finds no measurable difference between Metis Hybrid Olala and StandardOne 3B or Jevstral 8B.
  • Strong on the hard tier. On the 111 hard questions it matches StandardOne 8B (62 against 63). StandardOne 8B’s lead comes entirely from the “original” tier, where it gets 71 of 72 right and we get 59.
  • Decisive on yes/no questions. JevBench v1.6 only counts a yes/no answer if the model is at least 80% sure of the right side. Metis Hybrid Olala clears that bar on 60.5% of hard yes/no questions. StandardOne 8B clears it on 39.5%, Jevstral 8B on 28.9% and StandardOne 3B on 15.8%.

For context: TypeSafe’s hosted Jev, a closed US model, answered 205 of these questions in our test, and our own Qwen-based Metis Hybrid 4B answered 195. On this benchmark the European base costs us 26 questions. We expect that gap to shrink as European base models grow.

What we have to fix

Two numbers show where the work is.

Calibration. On the hard tier, its stated confidence is on average 0.147 away from its real accuracy (the expected calibration error). Our own target is 0.10, and StandardOne 8B reaches 0.089.

Speed. Olala mini uses about 1B parameters per token, so each decision should be cheap. In our test it was not fast: on the same H200 GPU it took a median 177 ms per decision, against 23 ms for StandardOne 3B, 29 ms for StandardOne 8B and 61 ms for Jevstral 8B. The size is not the problem: it runs through plain transformers with Olala’s own model code, while StandardOne runs on SGLang’s optimised kernels. OVHcloud’s vLLM build serves Olala for chat but not with a decision head, and that serving path is what would make each decision cheaper.

OVHcloud plans Olala v1.1 for the end of 2026, with a 128,000-token context. Longer context should help on the long-policy questions where every model in our test struggles.

Try it

The model, its pointer head, its calibrator and the loading code are all in GreenPT/metis-hybrid-olala-v1.0-mini. We tested both ways to use it from a clean machine with no Hugging Face token. Run it inside OVHcloud’s image ovhaillm/olala-ai-training:20261117, which has the compiled kernels Olala needs. The image starts offline, so set HF_HUB_OFFLINE before you import huggingface_hub:

import os, sys
os.environ["HF_HUB_OFFLINE"] = "0"
os.environ["IDV2_BACKBONE_REVISION"] = "0fde460f37d0fe2c390619bb5729c9d81dc89091"
from huggingface_hub import snapshot_download

path = snapshot_download("GreenPT/metis-hybrid-olala-v1.0-mini")
sys.path.insert(0, path)
from integrated_decisions_v2 import DecisionLM

model = DecisionLM.from_pretrained("OVHaiLLM/Olala-v1.0-mini-7A1B").to("cuda").eval()
model.load(path)
model.systemone({"state": "Refunds require a receipt. The customer has no receipt.",
                 "questions": {"refund": {"type": "noul", "instructions": "Is a refund allowed?"}}})

The model card also has the steps to run it as an HTTP server. For the other Metis models, read Meet Metis and Meet Metis Hybrid. If you want to use a European decision model in your own product, get in touch.

Frequently asked questions

What is a decision model?

A decision model reads a situation and a question with fixed options, then returns a probability for each option in one pass. It writes no text for the decision, so it is fast and its answers are easy to check and log.

What does "Hybrid" mean in Metis Hybrid?

One model does two jobs. A pointer head picks an option and returns calibrated probabilities, and the same model also writes text, so it can explain a decision or answer a chat message. Olala mini is called a hybrid for a different reason: it mixes Mamba-3 and attention layers.

Is Metis Hybrid Olala fully European?

Its base model, Olala mini, was designed in France by OVHcloud and trained from scratch on the JUPITER, Jean Zay and MeluXina supercomputers. Like many models, Olala reuses an existing tokenizer (Qwen's), which its NOTICE file credits. The decision training is ours.

How does it compare with Jev and Metis Hybrid 4B?

TypeSafe's hosted Jev, a closed US model, answered 205 of the same 231 JevBench questions in our test. Our Qwen-based Metis Hybrid 4B answered 195. Metis Hybrid Olala answered 169.

How do I run it?

Download GreenPT/metis-hybrid-olala-v1.0-mini from Hugging Face. The loading code and a /v1/systemone server ship in the repository. Run it inside OVHcloud's Olala image, which has the compiled kernels Olala needs. The model card has tested steps.

What licence does it have?

The adapter builds on Olala mini, which is Apache-2.0. The training data includes datasets with their own terms, such as the Yelp reviews, so check the model card before you use it beyond evaluation.

Back to all posts