This week, the word autonomy surfaces in places that are rarely considered together.
On Tuesday, Mistral AI unveiled its largest model to date and linked the promised release of its weights to the ability of organisations and public administrations to run artificial intelligence under their own policies. From 8 to 10 October, Athens is hosting a conference co-organised by a Department of Social Anthropology whose title explicitly invokes the ethics of autonomy. Meanwhile, Pula is hosting a gathering devoted to a third dimension of the same problem: what happens to cultural memory when machines begin to process, classify and reconstruct it.
The word is the same. The relationships it describes are different.
For Mistral, autonomy means giving an organisation greater control over the model it uses. In Athens, it refers to the collective capacity to make decisions about an infrastructure and manage the dependencies that make it possible. Pula adds another dimension: who has a say when archives, languages, images and cultural categories become part of computational memory.
In brief
- Paris, 6 October. Mistral opens the preview of Mistral Large 4. Its weights are due by the end of the month, but the licence under which they will be released remains unknown.
- Pula, 7–9 October. DARIAH-HR hosts Digital Humanities & Heritage: Cultural Memory vs. Machine Memory.
- Athens, 8–10 October. The first Digital Commons Conference brings together anthropology, science and technology studies (STS), computer science, design and Indigenous studies to examine infrastructures and autonomy.
Paris — Who can run the model?
On Tuesday, Mistral AI opened the public preview of Mistral Large 4 (ML4) and announced that it would release the model’s weights before the end of October.[1]
The updated announcement describes a multimodal mixture-of-experts model with one trillion parameters, 52 billion of them active. The main v26.10 model page specifies 1.05 trillion parameters in total and 52 billion active, while another versioned documentation page still gives a figure of 49 billion active parameters. This discrepancy, which persists across official sources for a preview model still undergoing training, means these figures cannot yet be treated as a fully settled specification.[1][2]
The context window extends to one million tokens. The model was trained from scratch using 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centres, and the public preview is served from that same infrastructure.[1] Until the weights are released, general access is via an API. The self-hosting option that Mistral presents as central to its proposition is not yet available.
Nor is the licence governing those weights known.
The precedent matters. In December 2025, Mistral released the Mistral 3 family, including Mistral Large 3, under Apache 2.0, a permissive licence allowing models to be used, modified and redistributed, including commercially.[3] Mistral already calls ML4 open-weight, but neither its announcement nor its technical documentation specifies the rights that will accompany the files when they are released.
Until then, open describes the direction announced by the company, rather than known terms of use.
ML4’s presentation repeatedly invokes sovereignty. Mistral says the model will be available through an end-to-end European deployment operated by the company, independent of other digital service providers and governed by European law. In cybersecurity, it explicitly links open weights and self-managed deployment to an organisation’s ability to work under its own policies.[1]
The infrastructure nevertheless introduces a significant material dependency. The data centres belong to Mistral; the GPUs are NVIDIA’s.
On Wednesday, writing about MareNostrum 5, we observed that European AI sovereignty also depends on who can access computing capacity. In September, we argued that open does not mean shared. Possessing a model’s weights enables activities that a closed API does not, but it does not provide the industrial infrastructure needed to produce the model.
Mistral also presents ML4 as the first outcome of a roadmap funded by its €3 billion Series D round, which the company describes as the largest ever raised by a European technology firm.[1] Here too, the corporate claim needs to be distinguished from its wider implications. A European company can own its data centres and operate under European jurisdiction while remaining dependent on an international technological supply chain.
The very idea of autonomy thus begins to look less absolute.
Pula — What does a machine remember?
Hundreds of kilometres away, the sixth international Digital Humanities & Heritage conference, organised by DARIAH-HR and other European institutions, concludes today in Pula under the pointed title Cultural Memory vs. Machine Memory.[4]
The organisers caution that the contrast is not intended as a simple separation between human memory and technological processing. Their approach centres on the relationships between interpretation and computational abstraction, contextual knowledge and automated classification, cultural responsibility and technological acceleration.
The questions they pose are concrete. What do digital systems remember? What do they forget? Which histories become visible, searchable and reusable? Which remain fragmented, marginal or poorly represented? What changes when cultural memory is encoded in datasets, metadata, prompts, models and automated outputs?[4]
Yesterday’s programme included one of the two keynote lectures, Jennifer Edmond of Trinity College Dublin, Professor of Digital Humanities, speaking on Cultural Memory, Machine Thinking, and Infrastructural Ethics. As this edition went to press, we had not found a recording, slides or a publicly available account of the lecture. We therefore confine our discussion to the conference’s published framing, which foregrounds the mediations between cultural memory, computational abstraction and infrastructural responsibility.[4]
Today’s programme addresses archives, algorithmic visibility and the politics of memory directly, before a closing lecture by Benedikt Perak of the University of Rijeka, Between Cultural Memory and Machine Memory: Five Levels of AI in Teaching.[4]
The discussion in Pula can also be applied to Mistral’s announcement.
Mistral says that a significant portion of ML4’s training data was multilingual, covering more than 160 languages, including all the official languages of the European Union.[1] We have not found a sufficiently detailed public breakdown of the specific texts, images and other materials making up that corpus.
The EU AI Act requires providers of general-purpose AI models to publish a sufficiently detailed summary of the content used for training. In July 2025, the European Commission adopted a common template for that information.[5][6]
This obligation matters because weights alone cannot answer a basic cultural question. We may inspect parameters without knowing which archives, selections, languages and exclusions contributed to producing them.
A machine’s memory is not simply a matter of how much it can store. It also depends on what was allowed to enter it.
Athens — Who governs the infrastructure?
The Digital Commons Conference: Infrastructures, Design and the Ethics of Autonomy is taking place in Athens from yesterday until tomorrow. It is co-organised by the Department of Social Anthropology at Panteion University and the Institute of Informatics and Telecommunications at the Demokritos National Centre for Scientific Research.[7]
Its approach deliberately moves beyond seeing digital commons as finished technological products. The organisers describe them as situated practices of autonomy, equitable access and collective care, bringing together anthropology, science and technology studies, computer science, human–computer interaction (HCI), design, feminist and Indigenous studies, and civic technology initiatives.[7]
Thursday’s opening programme placed Alberto Corsín Jiménez, a researcher at Spain’s National Research Council (CSIC), first, with Augural Commons: When Digital Infrastructures Inaugurate Territories They Do Not Yet Possess. The official programme confirms the scheduled time and announced contributions, but by the close of this edition we had not found a subsequent recording or report that would allow us to attribute statements to what was actually said.[8]
The abstract draws on twenty years of ethnographic work in Madrid, among free-culture collectives, the autonomous and neighbourhood assemblies of the 15M movement, and solidarity networks that emerged during the pandemic. Corsín recognises that the digital commons tradition has developed tools for studying licences, moderation, governance and technological experimentation. He also calls for attention to ways of making things common that do not yet exist as bounded domains. They operate at the edges, in transition, gradually building their own conditions of existence.[8]
This approach extends the picture we drew on Monday of the social sciences and AI in Spain. The anthropology of technology is not merely a method being incorporated into computational programmes. There is also a longstanding tradition of ethnographic research into free software, infrastructures, movements and forms of technological organisation.
The programme’s first panel then took the discussion towards community infrastructures.
Adrián López Angulo, of REDES A.C. and LocNet, examines telecommunications projects built and governed by rural and Indigenous communities in Latin America. His cases include Radio Tosepan Limakxtum and Wiki Katat in the Sierra Norte de Puebla, Mexico, as well as the Jxa’h Wejxia Casil network in Colombia’s Cauca region.[8]
The starting point is a set of familiar inequalities: unequal access, the extraction of minerals and data, connectivity models associated with surveillance, and the spread of disinformation. These experiences offer a different relationship with infrastructure. Communities do more than receive connectivity. They integrate technology into their own structures of decision-making, cultural reproduction and territorial care.
This is where a fundamental distinction emerges.
Paraskevi Zotali, a social anthropologist and engineer, places commons theory in dialogue with lo comunal and comunalidad, concepts developed in Oaxaca by Floriberto Díaz and Jaime Martínez Luna.[8]
Territory, assemblies, service, collective labour and ritual are part of an understanding of shared life that reaches beyond the joint management of a resource. Zotali asks how far the category commons can help us understand Indigenous telecommunications, and how far it might obscure historical forms of collective life that have their own concepts, relationships and cosmologies.
The distinction also matters for AI.
A downloadable file can be open. A collectively maintained resource can become a common. An infrastructure embedded in territorial and cultural ways of life entails a different relationship again.
Daniela Sant’Ana and Ana Delgado, of the University of Oslo’s TIK Centre, offer a warning from another territory. Their research on environmental data infrastructures and Sámi reindeer herding in Sweden examines how openness policies can generate new conflicts and reinforce processes of appropriation and extraction. Data produced about a territory, a culture or a way of life do not lose their social relationships when they become accessible.[8]
The panel ends with a formulation particularly useful for returning to Mistral.
Anthropologist Philipp Budka, of the University of Vienna, compares broadband and transport infrastructures governed by communities in northern Canada. His conception of sovereignty is relational. Community management does not eliminate dependency on external funding, regulation or technical suppliers. It builds the capacity to negotiate those dependencies over time.[8]
Those who govern an infrastructure also live with its consequences. Ownership therefore has an ethical, rather than merely legal, dimension.
Mistral promises deployment independent of other digital service providers. Budka describes an autonomy measured by the capacity to govern interdependencies that do not disappear.
Thursday’s programme also introduced a second political scale. A panel on India addressed access to academic knowledge through Sci-Hub, the political economy of major publishers, and representations of Aadhaar, the national digital identity system. Here, the commons are entangled with market relations, administrative classification and struggles over access to knowledge.[8]
The debate also extends beyond academia. The conference is collaborating with TRANSEUROPA 2026, a festival taking place in Athens under the theme Reclaiming the Cloud(s). On Saturday 10 October, the two initiatives will participate in a citizens’ assembly on AI and border control, examining surveillance, migration management and access to rights. The consequences of digital infrastructure become particularly visible when its decisions affect people’s mobility and lives.[13][14]
Open, common, communal
These three scenes help distinguish words that are often used interchangeably.
In Paris, openness refers to the announced possibility of obtaining a model’s weights and deploying it under one’s own policies. In Athens, a common entails relationships of governance, responsibility and care around an infrastructure. The Latin American experiences also introduce lo comunal, a form of organising collective life that does not necessarily begin with the existence of a shared resource.
Pula adds memory.
Governing an infrastructure also means deciding what it can record, which categories it uses and under what conditions others will be able to reuse what it contains.
This afternoon, the Athens programme continues along precisely these lines. The Digital Commons, AI & Collective Governance panel brings together work on AI openness, generative extractivism and public digital sovereignty. Another panel then addresses critical literacy, data governance, disinformation and collaboration between humans and AI.[8]
Among the contributors are Karen Soacha-Godoy, of the MINKA Citizen Observatory at ICM-CSIC and the Open University of Catalonia (UOC), and Jaume Piera, of the Institut de Ciències del Mar. They will present SEXTANT, a framework for analysing how seemingly technical decisions in a citizen-science platform — licences, APIs, authentication, privacy and default workflows — distribute control, access and decision-making power in practice.[8]
Tomorrow, the conference turns to heritage. Digital Cyclades will present a community cultural archive built using open tools and vision-and-language models. Alexandra Bounia and Despina Catapoti will then examine the fragile promise of applying the logic of the commons to AI, archaeology and heritage.[8]
That panel will take place after this Radar is published. We will follow it.
Another, much more concrete test will arrive at the end of the month. Mistral is due to release ML4’s weights, at which point we will know the licence under which they can actually be used.
Open, common and communal describe different relationships with an infrastructure. Their significance depends on who can govern it, who takes part in its decisions, which dependencies can be negotiated and who is answerable for the consequences.
References
[1] Mistral AI. Introducing Mistral Large 4. 6 October 2026. https://mistral.ai/news/mistral-large-4/
[2] Mistral AI. Mistral Large 4. Technical documentation, v26.10. https://docs.mistral.ai/models/mistral-large and https://docs.mistral.ai/models/mistral-large-4-0
[3] Mistral AI. Introducing Mistral 3. 2 December 2025. https://mistral.ai/news/mistral-3/
[4] DARIAH-HR. 6th International Conference Digital Humanities & Heritage 2026: Cultural Memory vs. Machine Memory. Conference programme and overview. https://dhh.dariah.hr/2026/programme/
[5] Regulation (EU) 2024/1689, Article 53. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
[6] European Commission. Explanatory Notice and Template for the Public Summary of Training Content for general-purpose AI models. 24 July 2025. https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models
[7] Digital Commons Conference. About the Conference. Athens, 8–10 October 2026. https://digicommons.org/
[8] Digital Commons Conference. Digital Commons: Infrastructures, Design and the Ethics of Autonomy. Programme. Accessed 8 October 2026. https://digicommons.org/programme/
[9] Corsín Jiménez, Alberto; Estalella, Adolfo. Free Culture and the City: Hackers, Commoners, and Neighbors in Madrid, 1997–2017. Cornell University Press, 2023. https://www.cornellpress.cornell.edu/book/9781501767203/free-culture-and-the-city/
[10] Díaz, Floriberto. Escrito. Comunalidad, energía viva del pensamiento mixe. UNAM, 2007. https://www.libros.unam.mx/gd-floriberto-diaz-escrito-comunalidad-energia-viva-del-pensamiento-mixe-ayuujktsen-9786070262081-libro.html
[11] Martínez Luna, Jaime. Eso que llaman comunalidad. Oaxaca, 2010. https://books.google.com/books/about/Eso_que_llaman_comunalidad.html?id=1v9ZMwEACAAJ
[12] Escobar, Arturo. Designs for the Pluriverse: Radical Interdependence, Autonomy, and the Making of Worlds. Duke University Press, 2018. https://www.dukeupress.edu/designs-for-the-pluriverse
[13] Digital Commons Conference. Partners — TRANSEUROPA 2026. https://digicommons.org/partners/
[14] TRANSEUROPA 2026. Programme — Digital Borders: a citizens assembly on Migration, border control and new technologies. https://transeuropafestival.eu/programme/