Tested configuration

Strix Halo × llama.cpp Vulkan × Qwen3.6-35B-A3B — strict JSON automation

Deterministic extraction, classification and structured generation in strict JSON: a perfect task-success rate in both reasoning modes on this corpus — with a seventeenfold end-to-end time difference between them.

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…)
Workloadstructured-agent-v1 @ 2026-08-03 — 16 deterministic JSON tasks, EN+RU
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

Does strict JSON output stay reliable for everyday automation tasks — extract fields from a message, classify a review, produce a structured plan — on this exact configuration?

Verdict: PASS WITH LIMITS. Every answer in every repeat parsed as strict JSON and matched the ground truth per key, in both reasoning modes. The limits are real and listed below.

What was measured

Sixteen deterministic tasks, EN and RU: extraction with an exact expected object (names, dates, negative temperatures, booleans — including Cyrillic values behind Latin keys), classification over closed label sets, and generation with a structural contract. The gate strips code fences, parses strictly, and compares required keys against the ground truth. A parse failure and a wrong value are distinguishable in the records, giving parse and task success separately.

Results

Task success with reasoning disabled: 100.0% of requestsrepeated — and with reasoning enabled at a 4096-token budget it is the same: 100.0% of requestsrepeated.

The difference is time. A complete valid answer takes a median of 844msrepeated without reasoning versus 14677msrepeated with it — on this corpus the reasoning pass bought no measurable correctness and cost roughly seventeenfold in wall-clock time.

The limits in “pass with limits”

Revalidation triggers

Runtime image change, model artifact change, workload revision change (especially the sandboxed agent loop), driver/kernel change. Harness and corpus are public: agmind-bench.

← Tested configurations