Strix Halo × llama.cpp Vulkan × Qwen3.6-35B-A3B — everyday assistant

The question this card answers

Is single-user streaming chat responsive on this exact configuration for everyday human tasks — messages, emails, explanations for a child, plans?

Verdict: PARTIAL. The configuration passes cleanly in one operating setting, fails structurally in another, and one quality gate never fully clears in any setting. The details are the product; here they are.

What passes

With the reasoning block disabled via the chat template, the first token of the answer arrives at a median of 210msunit_replicated — the answer starts with the stream, and the completion rate across six repeated runs on two units was 100.0% of requestsunit_replicated, with the language and repetition gates passing throughout.

What fails

With reasoning enabled at a 1024-token completion budget, 58.3% of requestsunit_replicated of everyday requests returned HTTP 200 with an empty answer: the reasoning pass consumed the whole budget before the reply began. An application checking status codes records these as successes.

Raising the budget to 4096 tokens removes the empty answers but the first answer token then arrives at a median of 20191msunit_replicated — tens of seconds of visible “thinking” before a reply to a request like “write a two-sentence message to a neighbor”.

Meanwhile the conventional time-to-first-token is 212msunit_replicated in all three settings — it cannot distinguish the passing configuration from the broken one. The full analysis is in the report Time-to-first-token does not describe a thinking model.

What never fully clears

The per-prompt word budget: in every operating setting a large share of answers exceeded the length the prompt asked for. That is a property of the model’s verbosity on this corpus, not of the reasoning setting, and it is why this card is not PASS WITH LIMITS.

Replicated on a second unit

All three single-user cells were re-run on a second, commercially identical GTR9 Pro against a byte-identical model artifact: the pattern reproduced (instant answers without reasoning, an answerless majority at the 1024 budget, tens of seconds at 4096), and the single-user numbers above are derived across both units. One detail the replication added: on the second unit, one request out of 48 exhausted even the 4096-token budget — a larger budget lowers the risk of an empty answer; it does not eliminate it.

Under concurrency (closed loop)

With reasoning disabled and 4 concurrent streams, the median first answer token moves to 338msrepeated; with 8 streams, to 865msrepeated, with completion at 100.0% of requestsrepeated and

respectively. This is closed-loop measurement (each virtual user waits for their answer before asking again) — it shows what each of N simultaneous users feels, and it is not an arrival-rate capacity or SLO claim.

Operating envelope implied by the evidence

For an everyday-assistant role on this configuration: disable the reasoning block, or budget at least 4096 completion tokens and accept the wait. Do not ship the default reasoning-on / 1024-token combination; treat empty-answer responses as errors in your monitoring.

What would change this verdict

Answer-quality evaluation of the no-reasoning mode (untested here — plausibly worse), concurrency beyond 8, and a runtime or model artifact update — any of these triggers revalidation. Raw evidence bundles and the harness are public: agmind-bench, methodology on the evidence & data page.