My bets on open models, mid-2026
Interconnects · Nathan Lambert · 2026-04-15
(No summary yet for this item — extraction summaries are still backfilling.)
Appears in
Extraction
Topics: open-weight-model-forecastschinese-ai-ecosystemai-market-dynamicsai-regulationreinforcement-learning-training
Claims
- Open models will not fully close the capability gap with closed models across all domains; the long-term trajectory is primarily an economics question rather than a capabilities one.
- Chinese open-weight labs will face funding difficulties as soon as late 2026, with capability trajectory divergence visible 3–9 months later.
- The RL-dominated training era has created the first clear technical area where closed labs can dominate open-weight models — real-world agentic deployment data for online RL.
- Bans on open models above compute thresholds are practically unenforceable because another sovereign entity will train and release them publicly.
- The U.S. will slowly regain ground in open model adoption metrics starting in early 2027, led by models like Gemma 4, Nvidia Nemotron, and Arcee AI.
Key quotes
It's surprising that the top closed models did not show a growing capability margin over open models, based on compute differences for training and research, especially in the second half of 2025 and through today.
The RL dominated training era has increased the relevance of distribution to real-world use-cases as a key factor in continued capabilities improvements. These are tasks where users directly use tools like Claude Code or Codex to solve problems in their job with agents. This is the first clear technical area that closed labs can dominate open-weight models on capabilities.
Recurring calls to ban certain types of open models will continue to come but are in practice impossible to implement.