SignalTech/AI
MLPerf Training Introduces Its First LLM Post-Training Benchmark
High-priority alert based on source significance and keyword relevance.
Context
What happened: MLPerf Training is adding its first LLM Post-Training benchmark, starting with the v6.1 submission round in October 2026, alongside existing pre-training benchmarks.
Key numbers: v6.1 round October 2026; workload under 1024 GB300 hours; Qwen 3.5 397B with 17B active parameters per token; 10 of 512 routed experts (2%); 700 training and 251
Why it matters: The benchmark measures time-to-train to a target accuracy, not just throughput, so submitters can tune systems while keeping model quality in check. It targets SWE task gains on an already post-trained open-weight model.
What changed
MLPerf Training is adding its first LLM Post-Training benchmark, starting with the v6.1 submission round in October 2026, alongside existing pre-training benchmarks.
Why it matters
The benchmark measures time-to-train to a target accuracy, not just throughput, so submitters can tune systems while keeping model quality in check. It targets SWE task gains on an already post-trained open-weight model.