Vol.014 — Don’t Marry the LLM.
Date: 2026-04-27 | Newsletter
Key Summary
DeepSeek’s V4 models landed this week: 1.6 trillion parameters, 1 million token context window, performance rivalling closed-source frontier models, trained for roughly $5.6 million, and released under a true MIT license with no user caps or geographic blocks. Simultaneously, Meta quietly shipped Muse Spark as fully proprietary — API-only, no weights — despite having built its entire AI brand around open-source Llama. The Open Source Initiative noted that Llama was never truly open source to begin with. Zenta reads neither as a values story but as a strategy story: every major lab is racing toward dominance and picking the road that gets there fastest. OpenAI lobbies for regulations only companies its size can afford. Meta needs to be the default standard like Android. DeepSeek uses true openness to bypass Western regulatory barriers for ASEAN, Africa, and Latin America entirely. The business owner’s takeaway from Functional AI Partners: stop agonizing over which LLM to choose. In early 2026 the performance gap between open and closed models is effectively zero on knowledge benchmarks. LLM inference costs have dropped roughly 10x annually since 2022. The question isn’t which model is smarter — they’re all smart enough. The question is whether your systems are designed to switch when the landscape changes again, as it will. Do it simple. Don’t marry the LLM. Design for flexibility.