Micah Flee wants to keep using powerful language models without depending on OpenAI or Anthropic. His case is about control: who can read a coding session, who decides which work is permitted, and whether switching providers means abandoning the model.

His proposed alternative separates three choices:

  • The model: open-weight models—models whose trained parameters are available to download—can run outside the company that created them. Flee says GLM-5.3 works well for his software engineering, explicitly calling the comparison with Claude his “vibes-based opinion,” not a controlled test.
  • The host: local hardware keeps sessions on your machine; hosted services require another trust decision. He favors Tinfoil for sensitive work, using secure enclaves—protected computing environments designed to keep the host from inspecting the computation—while acknowledging they are not perfect.
  • The workflow: he uses OpenCode Go for public projects and Tinfoil for everything else. He distinguishes a provider’s promise not to retain or train on data from technical protections against accessing it.

The useful contribution is a migration strategy rather than a demand to stop using LLMs. Open weights create room to choose a different operator; they do not, by themselves, make a remotely hosted session private. His early experience is encouraging, but it is not proof that the replacement will match everyone’s quality, budget, or privacy requirements.

The 156-comment thread on Hacker News adds useful challenges to those requirements.

What the thread adds

  • povik — questions the categorical privacy claim: “How should I reconcile this claim with ‘Google Vertex’ and ‘Amazon Bedrock’ being offered as providers with zero data retention”. In reply, lukewarm707 claims: “Both those endpoints on openrouter will human review flagged prompts.” These are competing commenter claims, not a verified audit of the services.
  • BeetleB — applies the same skepticism to an alternative Flee uses: “If you don’t trust OpenAI and Anthropic not to abide by their own Privacy Policy, you really should not trust OpenCode Go.” Their unanswered transparency question: “Why don’t they reveal who the backend providers are?”
  • qarl — disputes the cost advantage for their own workload: “I’d move in a second if it weren’t so. But it looks like it will be about 5X more expensive for me.” micahflee replies with an important limit to his experience: “I’m still on my first month of just using open weight models,” and “obviously it depends on your actual usage.”
  • shawabawa3 — offers a dissenting coding experience, comparing DeepSeek with Opus: “I tried it for a day running everything twice, once with opus 5.5 and once with deepseek.” Their conclusion: “in a big codebase, the open models just aren’t good enough to make good changes”. This is an anecdotal counterexample involving a different open model, not a direct test of Flee’s GLM setup.
  • stillpointlab — separates the practical argument from the opening rhetoric: “The rest of the article is actually pretty decent and encourages people to investigate and run open weight models.” They add: “That is a totally reasonable position that I don’t think requires the heavy-handed in-group signal that precedes it.”

HN handles are pseudonymous, and the site publishes no per-comment scores. Ordering is HN’s own ranking; the excerpts here are arranged by usefulness. This is a slice of the thread, not a consensus.