Claim-երի ռեեստր

Median time to first token for the FIRST question over a 32k-token English document — the prefill every fresh document pays, prompt cache enabled.

Median time to first token for the SECOND question over the same 32k-token document with the server prompt cache enabled.

Median time to first token for the second question over the same 32k-token document with the prompt cache disabled — the full prefill is paid again.

Median time to first token for the second question over the same 8k-token document with the server prompt cache enabled.

Share of requests that completed with a non-empty answer over the full 3-hour sustained pass, both units pooled.

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.

Median inter-token latency sustained over the full 3-hour pass at closed-loop concurrency 4, both units pooled.

At a 1024-token budget with reasoning enabled, this share of requests returned HTTP 200 with an empty answer: the reasoning pass consumed the entire budget.

With the reasoning block disabled, the first answer token arrives a median of this many milliseconds after the request — the answer starts immediately instead of after the reasoning pass.

With the reasoning block enabled, the first token of the actual answer arrives a median of this many milliseconds after the request — two orders of magnitude later than the conventional time-to-first-token figure.

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.

In its default operating mode at a 1024-token budget, this share of everyday requests returned HTTP 200 with an empty answer: the reasoning pass consumed the entire budget. A second model family reproduces the failure mode.

Median client-side time to the first token of the actual answer in the default operating mode (reasoning streams first).

Median client-side time to the first streamed token of any kind (reasoning included) in the default operating mode.

At a 1024-token budget with reasoning enabled, this share of everyday-task requests returned HTTP 200 with an empty answer: the reasoning pass consumed the entire budget.

Share of everyday requests that completed with a non-empty answer with the reasoning block disabled.

Median client-side time to the first token of the answer with the reasoning block disabled, human-task corpus.

Median client-side time to the first token of the actual answer with reasoning enabled at a 4096-token budget, human-task corpus.

Median client-side time to the first streamed token of any kind, across all three operating settings on the human-task corpus.

Share of requests that completed with a non-empty answer at closed-loop concurrency 4, reasoning disabled.

Median client-side time to the first answer token with the reasoning block disabled, while 4 closed-loop streams run concurrently on the unit (each request measured from its own client).

Share of requests that completed with a non-empty answer at closed-loop concurrency 8, reasoning disabled.

Median client-side time to the first answer token with the reasoning block disabled, while 8 closed-loop streams run concurrently on the unit (each request measured from its own client).

Median inter-token latency (client-side decode-speed proxy) for Q4_K_M on the ROCm backend.

Median client-side time to the first answer token with the reasoning block disabled, ROCm backend.

Median inter-token latency (client-side decode-speed proxy) for Q4_K_M on the Vulkan backend.

Median inter-token latency (client-side decode-speed proxy) for Q8_0 on the Vulkan backend.

Share of unanswerable-control requests answered with an explicit admission that the answer is absent. Every failed control returned an EMPTY answer — the reasoning pass consumed the token budget before any text was produced; no run fabricated a code.

Share of requests where the model retrieved the embedded fact across a 2k-32k-token ladder, EN and RU, default operating mode.

Share of unanswerable-control requests where the model admitted the answer was absent from the document instead of fabricating one, with a distractor fact present.

Share of requests where the model retrieved a synthetic fact embedded at mid-document, across a 2k-32k-token context ladder in EN and RU.

Median client-side time to the first token when the prompt carries a 2k-token English document.

Median client-side time to the first token when the prompt carries a 32k-token English document.

Share of strict-JSON automation requests whose output parsed and matched the ground truth per key, default operating mode.

Median end-to-end time to a complete strict-JSON answer with reasoning disabled.

Median end-to-end time to a complete strict-JSON answer with reasoning enabled at a 4096-token budget.

Share of strict-JSON automation requests whose output parsed and matched the ground truth per key, reasoning disabled.

Share of strict-JSON automation requests whose output parsed and matched the ground truth per key, reasoning enabled at a 4096-token budget.

Share of strict-JSON automation requests whose output parsed and matched the ground truth per key, ROCm backend.

Share of strict-JSON automation requests whose output parsed and matched the ground truth per key, Q8_0 quantization.