Mini was visibly less accurate vs the original Kronos demo at
shiyu-coder.github.io/Kronos-demo. Base is the next-tier model
in the NeoQuasar Kronos family.
The base model uses a different tokenizer (Kronos-Tokenizer-base)
and a shorter max_context (512 vs 2048) — all three settings
updated together to keep them in sync, since a model/tokenizer
mismatch silently produces garbage.
CPU latency goes up roughly 6x per sample. Caller currently sends
sample_count=10 which lands at 6-15 minutes on this Space's free
CPU — within browser fetch budget. Upgrade Space hardware to T4
small (~\$0.40/hr) for ~30s-1min predictions.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
