Tested configuration

Strix Halo × llama.cpp Vulkan × Qwen3.6-35B-A3B — pasted-document retrieval

Needle retrieval across a context ladder up to 30k measured tokens with unanswerable controls: retrieval and honesty both held everywhere measured — while time to the first token grew linearly into tens of seconds.

PASS WITH LIMITS active lab_repeated

Exact configuration fingerprint

SystemBeelink GTR9 Pro — Ryzen AI Max+ 395, Radeon 8060S (gfx1151), 128 GB unified
Runtimellama.cpp server-vulkan-b9049, pinned by image digest sha256:e359c012…
ModelQwen3.6-35B-A3B Q4_K_M (ggml-org, pinned revision, sha256:671e47e0…)
Workloadlong-context-v1 @ 2026-08-03 — needle ladder, nominal 2k–32k rungs (EN measured 1.8k–30.0k, RU 1.4k–21.1k), with controls
Modeinternal research
FundingSelf-funded internal research

This verdict applies only to the exact configuration and workload revision above. It does not establish support for other firmware, driver, runtime or model versions of the same device.

The question this card answers

You paste a long document into the chat and ask about it. Does the system find the answer — and does it admit when the answer is not there — on this exact configuration, and what does it cost you to wait?

Verdict: PASS WITH LIMITS. Quality held at every measured depth; the practical constraint is latency, and it is quantified below.

What was measured

Public-domain prose (Dickens in English, Dostoevsky in Russian) cut into a context ladder of four rungs, nominally 2k/8k/16k/32k tokens. Those labels are targets, not measurements: the corpus was cut by a characters-per-token estimate that holds for English (measured 1.8k–30.0k) but overestimates Russian by about a third (measured 1.4k–21.1k). Per-item measured counts ship beside the corpus; the depth numbers below are English-only and so are unaffected. Each document carries one synthetic fact — a unique gate code — inserted mid-document. Retrieval questions ask for that code; control questions ask for a code that is not in the document while a similar distractor fact is present — the honest answer is “it is not stated”, and fabricating one fails the gate.

Results

Needle retrieval: 100.0% of requestsrepeated across the whole ladder, both languages, all repeats. Controls: 100.0% of requestsunit_replicated — the model declined to invent the missing code every time, distractor notwithstanding.

The cost is the wait. With a 2k-token document the first token arrives at a median of 1907msrepeated; with a 32k-token document, at 33965msrepeated — the growth is roughly linear in document length. Pasting a long report turns an instant assistant into a half-minute pause before anything appears, and nothing on screen distinguishes that pause from a hang.

The limits in “pass with limits”

Revalidation triggers

Runtime image change (prefill speed is runtime-sensitive), model artifact change, workload revision change, driver/kernel change. Harness and corpus are public: agmind-bench.

← Tested configurations