Nezna-io

LeMo: China Unveils a New Open-Source Multimodal AI Model

By Julien Mercier

5 months ago

Échiquier ancien avec pièces en bois et appareil photo argentique posé dessus, lumière douce du soir, ambiance réfléchie et stratégique
Antique chessboard and analog camera – symbols of strategy and precise capture – Nezna/generated by IA
In short
  • LeMo: 8B multimodal open-source model from China
  • Claims competitive or superior scores vs Llama-3.1 8B
  • Native text + vision, trained on 6T tokens
  • Limits: hardware needs, lower maturity than Western leaders
  • Impact: AI democratization, but real-world usage questioned

Like chess, where every move sets up the next, AI progresses through calculated steps. LeMo, released in the past 24 hours, embodies this patient strategy.

This 8-billion-parameter model, fully open-source from China, natively handles vision and text. Published benchmarks show competitive results against Llama-3.1 8B across multimodal and text tasks.

Strengths lie in training efficiency (6 trillion tokens) and complete weight openness, enabling independent research. Yet, as in analog photography, raw precision is not enough: data quality, robustness and real-context adaptation remain key challenges.

Abstract representation of a multimodal AI model processing text and image together, sober futuristic editorial style
Abstract representation of a multimodal AI model processing text and image together, sober futuristic editorial style – Nezna/generated by IA

Western outlets (The Verge, Ars Technica) highlight the surprising competitiveness, while Asian media (Nikkei Asia, SCMP) emphasize China’s accelerating ecosystem. Russian and European reports stay cautious, noting lack of large-scale deployment and comprehensive independent verification so far.

The core strategic question: does LeMo truly serve human needs or is it mainly a geopolitical chess move? Open-source speeds innovation, but without concrete, evaluated usage, it risks remaining a paper promise.

FAQ

Does LeMo really outperform Llama-3.1?
On official benchmarks yes, but independent verification is still pending.

Can I use it right now?
Yes, weights are released, but multimodal inference requires strong hardware.

Why is Chinese open-source AI advancing so fast?
Massive funding, data access and strong catch-up motivation.