Sunday, 22 Nov 2026
  • The interview
  • Moroccans
  • Hardware test
  • Start-ups
  • My favorites
  • Advertisement
  • Contact
Business Club
Silicon Valley Maroc – le mag tech marocain
  • Real Estate
  • Militaire
  • AI
  • Cybersecurity
  • Tech
  • Expatriation
  • Industry
  • English
    • Français
    • العربية المغربية
    • English
    • Español
    • 中文 (中国)
  • Tourisme
  • Numérique
  • Business
  • Finance
  • Marketing
  • Apple
  • Claude
  • Google
  • Grok
  • OpenAI
  • USA
  • Europe
  • Afrique
  • Asie
  • Golfe
LIVE
  • ●
  • Casa
  • Rabat
  • Marrakech
  • Tanger
  • Agadir
  • Fès
  • Meknès
  • Oujda
  • Nador
  • Essaouira
  • Dakhla
  • Kenitra
  • Laâyoune
Redimensionnement de policeAa
Silicon Valley Maroc – le mag tech marocainSilicon Valley Maroc – le mag tech marocain
Rechercher
  • Real Estate
  • Militaire
  • AI
  • Cybersecurity
  • Tech
  • Expatriation
  • Industry
  • English
    • Français
    • العربية المغربية
    • English
    • Español
    • 中文 (中国)
Vous avez déjà un compte ? Se connecter
Suivez-nous
  • Dakhla
  • Casa
  • Marrakech
  • Tech
  • Rabat
  • Maroc
  • Plan du site
  • Contactez-nous
© 2026 - Colmar.tech
AIBusinessMorocco

A Moroccan startup launches an AI that learns from its mistakes

By observing enterprise artificial intelligence funding announcements in recent weeks, one question keeps coming up: will organizations continue to rent their AI capabilities, or will they start owning part of them?

Farid N.
Dernière mise à jour : 19 November 2026 18h43
Farid N.
Partager
Une startup marocaine lance une IA qui apprend de ses erreurs

The $30 million funding round announced on November 17, 2026 by Remanence, a Paris-based startup cofounded by Moroccan entrepreneur Hamza Benchekroun, speaks directly to this debate.

Sommaire
  • A team from French lab H Company
  • What the funding round will actually pay for
  • How the platform turns mistakes into progress
  • Why this matters beyond a simple funding round
  • What to take away at this stage

Remanence does not present itself as yet another frontier model lab. It offers a post-training platform that lets companies turn the execution traces of their AI agents — every successful task, every error, every human correction — into learning environments. These environments are then used to specialize open-weight models through reinforcement learning. At the end of the process, the company keeps ownership of the model weights and can keep its data on its own infrastructure.

A team from French lab H Company

The three cofounders know each other from working together at H Company, the French lab specializing in agents capable of operating a computer. Pierre-Louis Cedoz, now CEO of Remanence, was a founding researcher and then CTO there, after working on Gemini post-training at Google DeepMind. Hamza Benchekroun, CTO, led post-training at H Company and previously deployed reinforcement learning methods at InstaDeep. Axel Moyal, Chief Product Officer, led work on agents, evaluations and environment synthesis there.

At H Company, the trio led the post-training of the open-weight Holo models. According to EQT Ventures, in March 2026 these models set a new state of the art on the OSWorld-Verified benchmark, one of the most demanding for long-horizon agentic tasks, while remaining significantly cheaper than some proprietary models.

What the funding round will actually pay for

The $30 million round was co-led by EQT Ventures and FirstMark. Zero Prime, Emblem, Drysdale, Global Founders Capital and Yellow also participated, along with several business angels including Olivier Pomel (Datadog), Philippe Corrot (Mirakl), Romain Niccoli (Pigment), Synthesia cofounders Victor Riparbelli and Steffen Tjerrild, and Mehdi Ghissassi (AI71).

EQT Ventures

The funds will mainly go toward building a dedicated GPU cluster in Europe, hiring researchers and engineers (mostly in Paris, with an early presence in London and San Francisco), and expanding design partnerships with European and American companies. Remanence is already working with design partners in the legal, customer support and enterprise software sectors.

How the platform turns mistakes into progress

The technical principle is fairly simple to state, but harder to implement. The platform connects to agents already in production. It analyzes their execution traces: which tools were called, where the agent failed, what correction was made. From these traces, it automatically generates realistic training environments and evaluation mechanisms. An open-weight model is then post-trained with reinforcement learning on these environments. As new traces come in, the model can be retrained. The idea is that each use improves the next, without the company having to manually rebuild datasets or simulation environments every time its processes change.

According to the founders and their investors, this mechanism aims to reduce both inference costs (by producing more specialized and potentially lighter models) and dependence on proprietary model providers whose pricing and terms of use can change. The data remains under the company’s control, and both training and deployment can take place in a public cloud, a virtual private cloud or on-premises.

Why this matters beyond a simple funding round

As AI agents multiply model calls and, with them, token spending, many companies are finding that their costs are rising without performance improving proportionally on their specific use cases. The traces these agents generate often go unused. Remanence is betting that these traces are the most valuable raw material for building specialized intelligence that the company durably controls.

This approach is part of a broader movement. Open-weight models have narrowed the gap with closed models on many tasks, and a growing share of companies are experimenting with or deploying models they can host or adapt themselves. FirstMark’s Matt Turck has described this shift as a move toward a hybrid stack: closed models rented for some capabilities, and specialized models owned for others.

For European companies, particularly in regulated sectors, the ability to keep data and weights on their own infrastructure also addresses sovereignty and compliance requirements. It is not a universal solution, however. It requires a sufficient volume of quality traces, the ability to properly assess the model’s progress, and an organization ready to manage the retraining cycle.

The limitations remain real. Building learning environments that faithfully reflect real processes is technically difficult; this is precisely the bottleneck Remanence claims to automate. The scarcity of talent capable of running reinforcement-learning post-training at scale partly explains the valuation of this team. Finally, the major labs and other players in agent infrastructure are also developing adaptation capabilities and feedback loops. The competitive advantage will come down to the quality of the automation, the speed of iteration, and the ability to demonstrate measurable cost and reliability gains with early customers.

What to take away at this stage

Remanence’s funding round is significant both for its size as a seed round and for the profile of its founders. It illustrates a conviction shared by several investors: lasting value in enterprise AI lies not only in access to the best-performing models of the moment, but in the ability to turn everyday usage into continuous improvement under the company’s control. Whether this conviction translates into concrete economic results will depend on technical execution and real adoption in demanding production environments. For now, the design partners and the compute cluster are the next observable milestones.

FAQ

Is Remanence a Moroccan startup?
No. It is a Paris-based startup cofounded by Moroccan entrepreneur Hamza Benchekroun (CTO) with two French cofounders. It plans to hire mainly in Paris, with an early presence in London and San Francisco.

How does the system learn from mistakes?
It captures the execution traces of agents already deployed (successes, failures, corrections), turns them into training environments, and post-trains an open-weight model through reinforcement learning. The model can be retrained as new traces come in.

Does the company really keep control?
According to the company and its investors, yes: the company retains ownership of the model weights, and the data and training can stay within its own infrastructure.

What will the $30 million be used for?
Mainly to fund a GPU cluster in Europe, hire researchers and engineers, and develop design partnerships with companies in Europe and the United States.

Partager cet article
Whatsapp Whatsapp E-mail Copier le lien Imprimer
ParFarid N.
At a time when Morocco’s digital transformation is accelerating, protecting our digital assets has become an absolute national priority. As a cybersecurity expert, my mission is to secure Morocco’s digital space against emerging threats. I assist public and private organizations in building robust defense strategies capable of safeguarding our data sovereignty and ensuring the continuity of essential services.
Article précédent Morocco manufactures anti-drone nets
Article suivant Le Polisario algérien, pris en étau entre le Maroc et la Mauritanie The Algerian Polisario, caught between Morocco and Mauritania
Laisser un commentaire

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Silicon Valley

Soutenez notre média gratuit !
Don tech Maroc
maroc start-up morocco annuaire entreprise

Vous pourriez aussi aimer

Les espagnols fuient la ville de Ceuta au Maroc
MoroccoSpain

Spaniards are fleeing the city of Ceuta in Morocco

Par Julie - Sanchez
Les Marocains, premiers entrepreneurs étrangers individuels en Italie
BusinessItalyMorocco

Moroccans are the leading foreign sole proprietors in Italy

Par Azedine - Gh
Artillerie de roquettes : le Maroc creuse l’écart avec l’Espagne
MilitaryMoroccoSpain

Morocco widens its lead over Spain in rocket artillery

Par Foxtrot
tgv morocco al boraq maroc
MoroccoTransport

$10 billion to put all of Morocco on the rails

Par Ibtissam Harjiss
Silicon Valley Maroc – le mag tech marocain
Facebook X-twitter Rss Linkedin

A Propos

SiliconValley – le mag tech marocain se veut une plateforme indépendante gratuite dédiée à l’innovation, au numérique et aux nouvelles technologies au Maroc.

À la croisée de l’actualité tech internationale et des dynamiques locales, le magazine met en lumière les startups marocaines, les entrepreneurs, les talents, les innovations et les tendances qui façonnent l’écosystème tech national.

Analyses, décryptages, interviews et dossiers de fond : SiliconValley ambitionne d’informer, d’inspirer et de connecter une nouvelle génération tournée vers l’avenir, avec un regard moderne, critique et résolument marocain.

Categories

  • Dakhla
  • Casa
  • Marrakech
  • Tech
  • Rabat
  • Maroc
  • Plan du site
  • Contactez-nous

Liens Utiles

  • Bourse
  • Interview
  • MRE
  • Tests
  • Start-ups
  • Pub
  • Contact

Connectez-vous

Nom d'utilisateur ou adresse e-mail
Mot de passe

Mot de passe oublié ?