Qwen3.6-35B-A3B Q4_K_M on Ryzen AI Max+ 395 (llama.cpp Vulkan) — time to the first token of any output: 221 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 the first token of any output was 221 ms (median over all valid requests of three cells). The measurement was taken under the frozen interactive-assistant-v1@2026-08-02 workload; evidence level lab_repeated, 9 valid runs across 1 physical unit. The value is re-derived from the raw run records on every CI build.

Statement

On the tested Strix Halo configuration, the first streamed token of any kind arrives in a median of this many milliseconds — but on this model that first token is reasoning, not the answer.

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-v1@2026-08-02

Aggregation

median over all valid requests of three cells

Evidence level

lab_repeated

Status

active

Limitations

Nine runs across three cells on one unit — repeated, not unit-replicated. Measured at concurrency 1 on a quiesced node with a desktop session present. Says nothing about when a usable answer starts; see the answer-token claim.

Runs

• run-20260802-nothink-a
• run-20260802-nothink-a-r2
• run-20260802-nothink-a-r3
• run-20260802-budget4k-a
• run-20260802-budget4k-a-r2
• run-20260802-budget4k-a-r3
• run-20260802-gonogo-a2
• run-20260802-budget1k-a-r2
• run-20260802-budget1k-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/ttft-any-token.sql

Cite this claim

AGmind Systems Lab. "On the tested Strix Halo configuration, the first streamed token of any kind arrives in a median of this many milliseconds — but on this model that first token is reasoning, not the answer." Claim strix.qwen36.interactive.c1.ttft-any-token (221 ms), evidence level lab_repeated, scope interactive-assistant-v1@2026-08-02. https://agmind.ai/claims/strix.qwen36.interactive.c1.ttft-any-token/

@misc{agmind_strix_qwen36_interactive_c1_ttft_any_token,
  author       = {{AGmind Systems Lab}},
  title        = {On the tested Strix Halo configuration, the first streamed token of any kind arrives in a median of this many milliseconds — but on this model that first token is reasoning, not the answer.},
  howpublished = {\url{https://agmind.ai/claims/strix.qwen36.interactive.c1.ttft-any-token/}},
  note         = {Claim strix.qwen36.interactive.c1.ttft-any-token: 221 ms; evidence level lab\_repeated; scope interactive-assistant-v1@2026-08-02},
  year         = {2026}
}

Machine-readable: /claims/strix.qwen36.interactive.c1.ttft-any-token.json · full registry