Ministral Architecture001 — validation generation throughput

The 14B training recipe generated slowly, one reason faster architectures were studied.

Single reading

5.15 tokens/s

MEASURED recorded 21 September 2026

Protocol

Test definition

Mean tokens/s over the 26-case validation generation (mean 3,420 ms per case).

Limits

Limitations

  • Includes prefill of long contexts with CPU-hosted endpoints (memory recipe); not a serving benchmark.

Log

Recorded values

Date Value Evidence Samples Source
21 September 2026 5.15 tokens/s MEASURED 26 docs/ledger/FINAL_ARCHITECTURE_001_SELECTION_DISPOSITION.json

Source paths refer to the private JARVIS repository and runtime records. They are listed for audit and are not published.