Swift TP1 qualification on one RTX PRO 6000

Qualification accepted, November 1, 2026.

The tested SGLang/MTP3 recipe passed its retained protocol, context, image, agentic, SWE pilot, stability, capacity and resource gates, with the limitations below. This finding makes no public claim about current deployment, routing, aliases, or client assignment.

Measured serving envelope

The fresh final capacity population is ten one-repetition cells (72/72 requests), distinct from the earlier 60-cell matched historical comparison. The tested identity is ukisai/Swift-1.5-Qwen3.8-27b revision b4c84d42903a8646b25857eb2827288d92ed85a4, served by SGLang 0.5.19 image sha256:37bbbd3444732a464bbc68dee4fb0164e0ce9e18e2f027f3fc967f1152d3c262. It uses BF16 weights, FP8 E4M3 KV, FP32 SSM state, TP1, EAGLE MTP3 (top-k 1, four draft tokens), 262,144-token context, evaluated interactive concurrency two, and a 16,384-token output reserve. At C1, median decode was 63.44 tok/s at 8K and 62.46 tok/s at 32K. The selected C2 operating point improved aggregate output over C1, but its 32K median decode was 13.43 tok/s, below the 15 tok/s soft objective; first-output p95 was 16.69 s, within the 30 s ceiling. C1 remains the latency-first option. Final C16 is queued requested load rather than sixteen active requests: configured-derived engine active capacity is nine while final C16 sampled running maximum was eight, so it is not evidence of sixteen simultaneous active decodes. C1/C2/C4 cells have four requests, C8 has eight, and C16 has sixteen; sampled p95 is descriptive, not a stable production tail estimate.

The final resource observer reconciled 11,949 engine samples and 2,992 slow samples with zero OOM deltas, zero identity changes, and no use of the protected second GPU. Deep health passed after the capacity sweep.

Quality and retained negative evidence

The final fixed SWE population completed 5/5 official grades and resolved 4/5; one Django case remained unresolved. The retained strict formatting diagnostic failed 4/4 and the R5 resource rejection remains a negative outcome. The strict failure means ordinary-summary capacity numbers are not exact controlled-decode results. The historical matched no-spec/MTP comparison remains separate from the fresh final population; only C1--C8 were otherwise matched, and the MTP C16 effective-cap difference prevents a C16 cross-arm claim.

Evidence

Artifact manifest · Evidence index · Source registry · Final capacity · Resource validation · Historical comparison chart.

The bundle contains redacted qualification snapshots. Original file-byte hashes and derivative hashes remain distinct; native canonical receipt digests use their documented serialization. Operational deployment records are private.