# Qwen3.6-35B-A3B Q4_K_M на Ryzen AI Max+ 395 (llama.cpp Vulkan) — decode-pace drift over three hours, 4 одновременных запроса: 1.6 %

На системе Beelink GTR9 Pro — AMD Ryzen AI Max+ 395, Radeon 8060S (gfx1151), 128 GB LPDDR5X-8000 unified под управлением llama.cpp (server, Vulkan backend), b9049 (server_fingerprint b9049-2496f9c14) с моделью ggml-org/Qwen3.6-35B-A3B-GGUF @ baec3ebee244 (Q4_K_M) измеренная величина «decode-pace drift over three hours», 4 одновременных запроса составила 1.6 % (last window vs first window, pooled over both units). Измерение выполнено под замороженным ворклоадом endurance-30m-v1@2026-08-03; уровень доказательности — lab_single_run, 2 валидных прогонов на 2 физическом(-их) юните(-ах). Значение перевыводится из сырых записей прогонов на каждом CI-билде.

## Statement

> Relative change of the median inter-token latency between the first five minutes and minutes 175-180 of a continuous 3-hour closed-loop pass, across both units.

## Evidence

| Field | Value |
| --- | --- |
| Claim id | `strix.qwen36.endurance.c4.itl-drift-180m` |
| Value | 1.6 % |
| Metric | decode-pace drift over three hours |
| Aggregation | last window vs first window, pooled over both units |
| 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) |
| Workload scope | endurance-30m-v1@2026-08-03 |
| Evidence level | lab_single_run |
| Status | active |
| Runs | 2 across 2 unit(s) |
| Published | 2026-08-03 |
| Last changed | 2026-08-03 |

**Limitations.** One 3-hour pass per unit — the pattern reproduced on both commercially identical units, but repeats within a unit are pending. Closed-loop concurrency 4, reasoning disabled, this corpus — not a general thermals verdict. Ambient not instrumented; die temperatures documented in the run notes.

## Raw evidence

- Sealed run bundle `run-20260803-end-c4-a`: https://github.com/botAGI/agmind-lab/tree/main/runs/run-20260803-end-c4-a
- Sealed run bundle `run-20260803-end-c4-b`: https://github.com/botAGI/agmind-lab/tree/main/runs/run-20260803-end-c4-b
- Derivation query: https://github.com/botAGI/agmind-lab/blob/main/catalog/claims/sql/itl-drift-180m.sql
- Machine-readable: https://agmind.ai/claims/strix.qwen36.endurance.c4.itl-drift-180m.json · full registry https://agmind.ai/claims.json

## Cited on

- https://agmind.ai/ru/reports/beelink-gtr9-pro-local-llm/
- https://agmind.ai/ru/reports/local-llm-homelab-hardware/
- https://agmind.ai/ru/reports/how-many-tokens-per-second-is-enough/
- https://agmind.ai/ru/reports/strix-halo-sustained-load-3-hours/
- https://agmind.ai/ru/reports/what-128gb-unified-memory-runs/
- https://agmind.ai/ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-endurance/
- https://agmind.ai/ru/measured/qwen3-6-35b-a3b-q4-k-m-on-ryzen-ai-max-395/

## Cite this

AGmind Systems Lab. Qwen3.6-35B-A3B Q4_K_M on Ryzen AI Max+ 395 (llama.cpp Vulkan): decode-pace drift over three hours, 4 concurrent requests — 1.6 % (last window vs first window, pooled over both units; 2 runs on 2 units; evidence level lab_single_run; workload endurance-30m-v1@2026-08-03). Claim strix.qwen36.endurance.c4.itl-drift-180m. https://agmind.ai/claims/strix.qwen36.endurance.c4.itl-drift-180m/

Registry changes feed: https://agmind.ai/claims/changes.xml · Corrections to published results: https://agmind.ai/errata/ · Data licensed CC BY 4.0.
Source page: https://agmind.ai/ru/claims/strix.qwen36.endurance.c4.itl-drift-180m/