15.17 Open versus closed weights
The frontier has split into two ecosystems: closed-weights labs (OpenAI, Anthropic, Google DeepMind, xAI) and open-weights labs (Meta, DeepSeek, Mistral, Alibaba's Qwen team, AI2). The split has been one of the dominant policy and economic stories of 2024–2026.
The argument for open weights
Open weights enable: independent safety research (mechanistic interpretability researchers cannot work on closed models without access); reproducibility (closed-weight benchmark numbers are unverifiable); democratisation (small companies and individuals cannot fine-tune closed models for their needs); resilience (a single lab's outage or business decision does not break the ecosystem); deployment in air-gapped or sovereign environments.
The argument against
Open weights enable: removal of safety fine-tuning (any RLHF-instilled refusals can be undone with modest fine-tuning); use by malicious actors (bioweapon synthesis assistance, disinformation campaigns); proliferation of capabilities to actors who would be denied API access; copyright violations (models trained on copyrighted data cannot be retracted).
The DeepSeek moment
In January 2025, DeepSeek-V3 and DeepSeek-R1 were released openly with full methods papers. The performance was at or near the closed frontier. The training cost, ~$5.6 M for V3, was a small fraction of what closed labs were spending. The combination of open weights, frontier capability, and dramatically lower training cost briefly moved equity markets and substantially altered the strategic calculus. Two consequences:
- The closed-frontier premium narrowed. By April 2026 the best open-weights models (DeepSeek-V3 and R2, Qwen3, Llama 4) are within a few percentage points on standard benchmarks of the best closed-weights models, at a fraction of the inference cost.
- The capabilities-proliferation question became unavoidable. By April 2026 the US, EU and UK have all promulgated some form of frontier-AI regulation; the EU AI Act came into full force in 2025.
The split is unlikely to resolve cleanly. In the short run, the closed frontier remains marginally ahead on the hardest reasoning tasks. In the medium run, the pace of open-source replication (typically 3–9 months) is fast enough that "frontier" and "open" are largely overlapping circles.