My Knowledge Base

Open-Source AI

3 min read

Open-Source AI

Open-source AI refers to the practice of publicly releasing model weights, training datasets, and/or training code — as opposed to serving AI exclusively through proprietary APIs. The movement has accelerated since 2023 with releases from Meta (Llama), nvidia (Nemotron), and Chinese labs including deepseek and MiniMax.

NVIDIA’s Strategy

jensen-huang articulates three reasons NVIDIA pursues open-source AI (Lex Fridman Podcast #494, 2026):

  1. Co-design research — Building and releasing models gives NVIDIA hands-on experience with evolving architectures, informing hardware design 2–3 years in advance
  2. AI diffusion — Proprietary-only AI excludes researchers, smaller countries, and specialised industries; open weights enable every sector to participate in the AI revolution
  3. Multimodal coverage — AI is not just language; domain-specific models (biology, physics, weather, robotics) need to stay at the frontier; NVIDIA publishes these to serve industries that cannot build their own

Nemotron 3 Super (120B-parameter MoE model, open weights + data + training recipe) is NVIDIA’s flagship open-source release as of 2026.

China and Open Source

Huang observes that Chinese AI labs contribute disproportionately to open source because of a cultural “schoolmate network”: engineers at competing firms are friends and former classmates who share knowledge freely. “What are we protecting?” — their social graph already spans competitors. This produces rapid innovation cycles and explains the strength of deepseek and similar labs.

Tension with Proprietary Models

Huang acknowledges the tension: world-class proprietary models (GPT-4, Claude) deserve to be products; open-source models enable research and innovation on top of and around them. Both coexist. Open-source models also serve as a hedge: if proprietary APIs become gatekeeping bottlenecks, open weights ensure no single vendor can block AI adoption.

The Chinese Open-Weight Advantage (2025–2026)

nathan-lambert and sebastian-raschka add important texture to the open-weight landscape (fridman-lambert-raschka-2026-state-of-ai). Chinese AI labs — deepseek, MiniMax, Kimi (Moonshot), Z.ai — trend toward:

  • Large open-weight MoEs with permissive licences (fewer restrictions than Meta’s Llama user-cap terms)
  • Attractive for enterprise fine-tuning on proprietary domain data without vendor API lock-in
  • A signal to global developers that Chinese compute doesn’t require API-level trust

The structural motivation: where API security concerns block use of Chinese-hosted inference (government, financial, healthcare), open weights let enterprises self-host without dependence on any Chinese-controlled endpoint.

deepseek R1 (January 2025) was the geopolitical AI moment of 2025 — near-SOTA performance at a fraction of claimed training cost — and triggered a wave of Chinese open-weight releases. By early 2026, deepseek was losing its open-weight crown to Z.ai (GLM), MiniMax, and Kimi K2 Thinking.

A landmark development: anthropic lost a $1.5B lawsuit in 2026 for torrenting (rather than purchasing) books for training data. This signals that training data sourcing decisions carry real legal liability and that courts are beginning to treat large-scale data acquisition methods as legally consequential. Labs are expected to shift toward licensed corpora and synthetic-data pipelines. See anthropic.

Sources: fridman-huang-2026-nvidia-ai-revolution | fridman-lambert-raschka-2026-state-of-ai

Graph View

Backlinks