Baseline
The production model is scored on an eval suite synthesized from its own failures, plus your regression suites.
Lab · Benchmarks
We measure what your model fails at — not generic leaderboards.
Pacer benchmarks every training cycle against held-out production failures and your own regression suites, so “better” always means better at your work.
FIG. 01 / EFFICIENCY
3.2×
Avg. improvement per cycle
FIG. 02 / VELOCITY
48 h
First improvement deployed
FIG. 03 / SCALE
100B+
Parameters on-stack
FIG. 04 / LABOR
0
Reward functions written
Self-generating eval suites
Auto-instrumented agents
RL across 2,500+ environments
Serverless model inference
Real-time training observability
Production data flywheel
Cycle
RL post-training runs against the failure surface across open and private environments.
Gate
A candidate only ships if it beats production on held-out failures without regressing any suite.
Report
Every cycle reports gain on held-out evals, with full lineage of environments and reward specs.
Gains compound every cycle.
Each cycle trains on the failures of the last one. A typical curve rises fast, then keeps climbing as the failure surface narrows.
| Suite | Domain | Task | Verifier |
|---|---|---|---|
| swe-repair | Code | Fix failing tests in real repositories | Tests passing |
| deepcode-verify | Code | Competitive programming tasks | Hidden-test accept |
| browser-ops | Agents | Multi-step web tasks | DOM-state match |
| sql-analyst-bench | Data | Text-to-SQL on messy warehouse schemas | Result-set match |
| i3-math | Science | Symbolic math problems | Symbolic match |
| support-agent (private) | Tool use | Synthesized from failed production traces | Auto-rubric |
Per-customer results are shared under NDA. Request a benchmark report for your model and workload.
[ Benchmark your model ]
See how much your model improves in one cycle.
Engineers, at full stride. Point PI at your checkpoint — first improvement in about 48 hours.