Qwen3.6-35B-A3B Q4_K_M on Ryzen AI Max+ 395 (llama.cpp Vulkan) — time to first answer token, 95th percentile, 8 concurrent requests: 1088 ms

The measurement

On a Beelink GTR9 Pro — AMD Ryzen AI Max+ 395, Radeon 8060S (gfx1151), 128 GB LPDDR5X-8000 unified running llama.cpp (server, Vulkan backend), b9049 (server_fingerprint b9049-2496f9c14) with ggml-org/Qwen3.6-35B-A3B-GGUF @ baec3ebee244 (Q4_K_M), the measured time to first answer token, 95th percentile, 8 concurrent requests was 1088 ms (95th percentile over valid requests, interpolated). The measurement was taken under the frozen interactive-assistant-v2@2026-08-02 workload; evidence level lab_repeated, 3 valid runs across 1 physical unit. The value is re-derived from the raw run records on every CI build.

Statement

95th-percentile client-side time to the first answer token with the reasoning block disabled, while 8 closed-loop streams run concurrently on the unit (each request measured from its own client) — the slow tail of the same requests behind the median claim.

Evidence

System

Beelink GTR9 Pro — AMD Ryzen AI Max+ 395, Radeon 8060S (gfx1151), 128 GB LPDDR5X-8000 unified

Runtime

llama.cpp (server, Vulkan backend), b9049 (server_fingerprint b9049-2496f9c14)

Model artifact

ggml-org/Qwen3.6-35B-A3B-GGUF @ baec3ebee244 (Q4_K_M)

Scope

interactive-assistant-v2@2026-08-02

Aggregation

95th percentile over valid requests, interpolated

Evidence level

lab_repeated

Status

active

Published

September 2, 2026

Limitations

Three repeated runs on one unit. Closed-loop concurrency 8 — not an arrival-rate/SLO capacity claim. Context split across server slots (8192 per slot). Each run has 16 corpus items, so even pooled across the listed runs the 95th percentile rests on a handful of the slowest requests — a coarse tail statistic, not a tail-latency model. Failed requests carry no answer-token time and are outside this statistic; the completion-share claim on the same runs accounts for them.

Runs

• run-20260803-v2-c8-nothink-a
• run-20260803-v2-c8-nothink-a-r2
• run-20260803-v2-c8-nothink-a-r3

Derivation

The value is re-derived from the raw run records by this query on every CI build — a published number cannot silently drift from its evidence.

catalog/claims/sql/ttfa-p95-nothink.sql

Cited on

Pages on this site that render this number:

Cite this claim

AGmind Systems Lab. Qwen3.6-35B-A3B Q4_K_M on Ryzen AI Max+ 395 (llama.cpp Vulkan): time to first answer token, 95th percentile, 8 concurrent requests — 1088 ms (95th percentile over valid requests, interpolated; 3 runs on 1 unit; evidence level lab_repeated; workload interactive-assistant-v2@2026-08-02). Claim strix.qwen36.interactive2.c8.ttfa-p95-nothink. https://agmind.ai/claims/strix.qwen36.interactive2.c8.ttfa-p95-nothink/

@misc{agmind_strix_qwen36_interactive2_c8_ttfa_p95_nothink,
  author       = {{AGmind Systems Lab}},
  title        = {95th-percentile client-side time to the first answer token with the reasoning block disabled, while 8 closed-loop streams run concurrently on the unit (each request measured from its own client) — the slow tail of the same requests behind the median claim.},
  howpublished = {\url{https://agmind.ai/claims/strix.qwen36.interactive2.c8.ttfa-p95-nothink/}},
  note         = {Claim strix.qwen36.interactive2.c8.ttfa-p95-nothink: 1088 ms; evidence level lab\_repeated; scope interactive-assistant-v2@2026-08-02},
  year         = {2026}
}

Machine-readable: /claims/strix.qwen36.interactive2.c8.ttfa-p95-nothink.json · BibTeX · CSL-JSON · full registry · changes feed