NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

From Lisbon to Laos: ModelBest CEO Li Dahai's Case for Open-Source AI as Development Infrastructure

At Web Summit 2026 in Lisbon, ModelBest (面壁智能) CEO Li Dahai told Xinhua that open-source AI lets developing countries acquire advanced capabilities at low cost, pointing to his company's role in training Laos' national large model. His argument landed amid a summit-wide display of Chinese open models entering European corporate production systems.

2026-11-18 · 875 words · NeuroAI
From Lisbon to Laos: ModelBest CEO Li Dahai's Case for Open-Source AI as Development Infrastructure

The pitch for open-source artificial intelligence usually runs through cost tables and benchmark charts. Li Dahai's version starts with a country that lacks the compute to train anything.

"Open-source AI can let more countries and enterprises around the world access advanced AI capabilities at lower cost, especially helping developing countries take part in the new round of technological change," said Li, chief executive of Chinese startup ModelBest (面壁智能), in an interview with Xinhua at the Web Summit 2026 in Lisbon, per the news agency's November 12 dispatch.

His concrete example sits well outside the usual AI conference circuit: ModelBest has participated, together with domestic partners, in building Laos' national large model, undertaking the model training work. For countries without the compute budgets or talent pools to train hundred-billion-parameter models from scratch, Li argued, open models mean a nation can still have AI infrastructure grounded in its own language, data and regulations — without starting from zero.

What ModelBest is and where it sits

ModelBest, headquartered in Beijing, is known for its MiniCPM family of small, efficient open models designed to run on end devices — phones, cars, offline industrial equipment. Per the company's own conference release, MiniCPM series models have surpassed 70 million cumulative downloads worldwide, and the company's automotive cockpit solution ships in seven production vehicle models from Chinese automakers. Li also spoke on the summit's main stage and at its AI Summit and China-Europe forum sessions, arguing that model capability per unit of compute — what his team calls the "densing law" — matters more than raw parameter counts for global adoption.

The Laos project is the clearest international expression of that thesis: rather than exporting a closed API, the company transfers an open model into a national context and trains it on local language and regulatory needs.

The summit backdrop: Chinese open models in production

Li's argument did not land in a vacuum. The same Xinhua dispatch, corroborated by AFP reporting ahead of the summit, documented how far Chinese open models have moved into European corporate systems:

  • Web Summit co-founder Paddy Cosgrave said from the opening stage that UK fintech firm Revolution (雷沃卢特) had replaced its OpenAI and Anthropic coding assistants with Zhipu AI's (智谱) open-source GLM model series. "The story of Revolution is a window into the future of tech — and that future increasingly has a Chinese flavor," he said, per Xinhua.
  • Revolution's AI infrastructure lead, David Arbitbol, told Xinhua the GLM model is now the default coding model for the company's roughly 2,000 engineers, that average coding time has fallen about 60%, and that AI-related invocation costs have dropped nearly 90%.
  • French telecom operator Orange, automaker Renault and Germany's Siemens have publicly said they use Chinese models such as DeepSeek and Qwen (通义千问) as complements to American and European models, per the AFP account carried in the dispatch.
  • A key driver cited by European firms: open weights can be deployed on companies' own servers, preserving data control and capping costs — attributes that closed frontier-model APIs do not offer.

Web Summit 2026, held November 9–12, drew about 70,000 attendees, and organizers said Chinese participation reached a record scale this year, with the China pavilion appearing on many European attendees' must-visit lists.

Why the developing-country framing matters

The Laos example highlights a division in the global AI market that gets less attention than the US-China rivalry: the gap between countries that can afford frontier compute and those that cannot. For the latter group, the choice is not "closed American model versus open Chinese model" — it is often "open model or no model." Open weights, freely downloadable and modifiable, let a national government or a local telecom stand up domestic-language capability without the capital expenditure of pretraining.

That framing also carries commercial logic for the vendors. Every national model built on an open base creates a services, fine-tuning and support relationship that an API subscription does not — and it seeds the developer ecosystem that makes an open model family durable.

Honest limitations

  • Li Dahai's remarks and the Laos project description come from a Xinhua interview and the company's own conference communications; no independent documentation of the Laos national model's scope, funding or performance has been reviewed here.
  • The Revolution coding-productivity figures — 60% faster coding, nearly 90% lower AI call costs — are single-company internal estimates reported by its own infrastructure lead to Xinhua, not audited measurements, and definitions of "coding time" were not provided.
  • Download counts for MiniCPM are company-reported and cannot be independently verified.
  • The characterization of European corporate usage (Orange, Renault, Siemens) rests on AFP reporting relayed via Xinhua and public corporate statements; the depth of each deployment was not independently confirmed.

What readers can do now

  • Developers evaluating open models for non-English or low-resource languages can benchmark the GLM, DeepSeek, Qwen and MiniCPM families against their closed alternatives, since the price-performance claims above are company-supplied.
  • Policy teams in developing countries can study the Laos arrangement as a template — open-weight base model plus local training partnership — when weighing how to build domestic AI capability without frontier-scale compute.
  • Analysts tracking open-vs-closed dynamics can watch whether more European enterprises follow Revolution's pattern of switching default coding models, which would test whether the reported cost savings generalize.

Related coverage

Cite this article

NeuroAI (脑机网), "From Lisbon to Laos: ModelBest CEO Li Dahai's Case for Open-Source AI as Development Infrastructure," 18 November 2026, https://neuroai.site/a/na-people-li-dahai-mianbi-open-source-laos

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Builders” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories