Priority metric
Win condition
>=10% GPU-hour reduction at SLA
Scoring rule
Tier weights: 2k: 25% · 4k: 25% · 8k: 25% · 16k: 25%
Failure penalty: 0.1 × failed eligible fraction
Any scored request failure caps speed credit at zero before the deduction.
Correctness
Precision
Context window
Baseline repo
Baseline commit
Patch visibility
Manifest hash
Base image
Baseline engine
Workload
32 prompts per round from zai-org/LongWriter-6k
default/train · 6,000 rows · 5,120 max output tokens · Natural EOS · 90% of baseline minimum
C32 request concurrency · All four tiers overlap
Allowed paths
Denied paths
Target GPUs
B300
8× GPU per bench
Submissions
28
22 passing · 3 rejected
Scoring rule
weighted tier completion speedup
B300
p99 latency
2000 / 50 ms
TTFT / ITL ceiling