Model Status · EchoNest 3
EchoNest 3 is live.
Production now serves EchoNest 3, a ~56M-parameter transformer language base (promoted Nov 2, 2026), replacing EchoNest v2 (12M). Held-out story perplexity 4.57 vs v2’s 9.92. Trained CPU-only on ~246M tokens. A chat / code / tool-use fine-tune is in progress (not live yet), ETA around Nov 3.
Training Phases
01
1.26B experiment
Earlier from-scratch 1.26B mixed-data run. Finished and retired — it did not beat production.
02
Language base · EchoNest 3
~56M transformer trained CPU-only on ~246M tokens of story text.
03
Eval & promote
Beat EchoNest v2 on the same held-out story text (perplexity 4.57 vs 9.92). Promoted to production Nov 2, 2026.
04
Chat / code / tool-use fine-tune
Short fine-tune on top of the EchoNest 3 language base. In progress, not live yet; ETA around Nov 3.
Overall Progress
EchoNest 3 · language base
—
Current Metrics
Train Loss
—
current batch
Val Loss
—
last epoch end
Val Perplexity
—
lower is better
Grad Norm
—
gradient L2 norm
Learning Rate
—
current LR
Global Step
—
optimizer steps
Loss History
Batch training loss (last 200 batches)
no data yet
Latest samples
Waiting for generation probes…
—