What Veil was
Veil 1.1 is a fine-tuned Llama 3.1 3B — small, fast, and free with an Axion account. It was our first model, built before Lumen existed, and it's the reason Lumen exists at all: everything we learned fine-tuning and shipping Veil shaped how we built Lumen 1.2.5.
Why it's going away
We recently moved our model hosting off shared, rate-limited infrastructure and onto our own servers — a chance to take a hard look at what we actually run day to day. Lumen has grown into a faster, more capable model that covers everything Veil did and more, with far better safety alignment behind it. Running two models split our attention between infrastructure, safety testing, and improvements — attention that's better spent making one model genuinely good than keeping two adequate ones alive.
This isn't a comment on Veil's quality for what it was built to do. It's a decision to focus.
What happens, and when
- Jul 25, 2026Veil is marked deprecated, and goes offline earlier than planned — the compute it ran on is committed to the Project Crucible 600M rerun. We originally said it would keep answering requests through this period; it didn't, and that was our miss.
- Jul 30, 2026Veil comes back online and answers requests as normal, giving a real migration window rather than the one we promised and couldn't deliver.
- Aug 17, 2026Veil is fully retired. Its model card stays up on the Models page permanently — a memory of where Axion started, not an active endpoint.
What to use instead
If you were calling Veil directly, switch to Lumen (model: "lumen") — same OpenAI-compatible endpoint and the same free Axion account. Veil is unavailable right now, so if you are still pointing at it, moving today is the fix rather than waiting for August 17.