A matrix run measures prompt reading (prefill) and generation (decode) speed in one go, across several context lengths and numbers of simultaneous requests.
DeepSeek-V4.1-Flash-1017RTX PRO 6000 Blackwell Server Edition 768GB ×8 · vllm 0.32.0 · MXFP4+FP88 contexts × 8 concurrency levels
Peak decode
4,156.3 tok/s
Peak prefill
15.9K tok/s
DeepSeek-V4.1-Flash-1017RTX PRO 6000 Blackwell Server Edition 384GB ×4 · vllm 0.32.0 · MXFP4+FP88 contexts × 8 concurrency levels
Peak decode
3,724.7 tok/s
Peak prefill
17.9K tok/s
DeepSeek-V4.1-Flash-1017RTX PRO 6000 Blackwell Server Edition 768GB ×8 · vllm 0.32.0 · MXFP4+FP82 contexts × 3 concurrency levels
Peak decode
1,929.4 tok/s
Peak prefill
—
DeepSeek-V4.1-Flash-1017RTX PRO 6000 Blackwell Server Edition 384GB ×4 · vllm 0.32.0 · MXFP4+FP82 contexts × 3 concurrency levels
Sampling audited from every vLLM request: temperature 1, top_p 1, top_k 0, min_p 0, presence/frequency penalties 0, repetition penalty 1, max_tokens 262144; context explicitly 1048576. Two warmups and 5 timed samples, current immutable container, no new tuning. Arithmetic and four-part completion checked for all 7 answers; the battery question permits pooled (13, $5.80) or equal-share (7, $43.90) interpretations, so this is not a formal accuracy score. Preview shows final-answer excerpts; native outputText, SHA256 and metrics retained. Token rates include reasoning and exclude prefill. These checks do not establish general quality or production readiness.
Sampling audited from every vLLM request: temperature 1, top_p 1, top_k 0, min_p 0, presence/frequency penalties 0, repetition penalty 1, max_tokens 262144; context explicitly 1048576. Two warmups and 5 timed samples, current immutable container, no new tuning. All 7 generated programs, including all 5 timed samples, passed 8 isolated stdlib-only functional cases each. Preview shows final-answer excerpts; native outputText, SHA256 and metrics retained. Token rates include reasoning and exclude prefill. These checks do not establish general quality or production readiness.
Sampling audited from every vLLM request: temperature 1, top_p 1, top_k 0, min_p 0, presence/frequency penalties 0, repetition penalty 1, max_tokens 262144; context explicitly 1048576. Two warmups and 5 timed samples, current immutable container, no new tuning. Arithmetic and four-part completion checked for all 7 answers; the battery question permits pooled (13, $5.80) or equal-share (7, $43.90) interpretations, so this is not a formal accuracy score. Preview shows final-answer excerpts; native outputText, SHA256 and metrics retained. Token rates include reasoning and exclude prefill. These checks do not establish general quality or production readiness.
Sampling audited from every vLLM request: temperature 1, top_p 1, top_k 0, min_p 0, presence/frequency penalties 0, repetition penalty 1, max_tokens 262144; context explicitly 1048576. Two warmups and 5 timed samples, current immutable container, no new tuning. All 7 generated programs, including all 5 timed samples, passed 8 isolated stdlib-only functional cases each. Preview shows final-answer excerpts; native outputText, SHA256 and metrics retained. Token rates include reasoning and exclude prefill. These checks do not establish general quality or production readiness.
Temperature 1, top_p 1, maximum output 262144 tokens; code-v1, 2 warmups, 5 timed samples. All 7 generated programs, including all 5 timed samples, completed and passed eight isolated stdlib-only functional checks each: malformed input, since filtering, percentiles, timezone offsets, empty input, invalid record schema, missing latency, singleton percentiles. Preview shows the final answer only; full raw output and SHA256 retained, and token rates include reasoning. These checks do not establish general model quality or production readiness. Different protocol from the earlier temperature-0 capped runs, which are flagged for quality investigation; no percentage comparison is claimed.
Temperature 1, top_p 1, maximum output 262144 tokens; code-v1, 2 warmups, 5 timed samples. All 7 generated programs, including all 5 timed samples, completed and passed eight isolated stdlib-only functional checks each: malformed input, since filtering, percentiles, timezone offsets, empty input, invalid record schema, missing latency, singleton percentiles. Preview shows the final answer only; full raw output and SHA256 retained, and token rates include reasoning. These checks do not establish general model quality or production readiness. Different protocol from the earlier temperature-0 capped runs, which are flagged for quality investigation; no percentage comparison is claimed.
QUALITY INVESTIGATION: this temperature-0, 8192-token canonical run includes unfinished reasoning; the 8-GPU preview contains a pathological repeated-sentence loop. These token rates are not validated useful-answer throughput or evidence of production readiness. Matched current/reference, DSpark/autoregressive and sampling controls are running; do not use this result to infer model quality.
QUALITY INVESTIGATION: this temperature-0, 8192-token canonical run includes unfinished reasoning; the 8-GPU preview contains a pathological repeated-sentence loop. These token rates are not validated useful-answer throughput or evidence of production readiness. Matched current/reference, DSpark/autoregressive and sampling controls are running; do not use this result to infer model quality.
Sampling audited from every vLLM request: temperature 1, top_p 1, top_k 0, min_p 0, presence/frequency penalties 0, repetition penalty 1, max_tokens 262144; context explicitly 1048576. Two warmups and 5 timed samples, current immutable container, no new tuning. Arithmetic and four-part completion checked for all 7 answers; the battery question permits pooled (13, $5.80) or equal-share (7, $43.90) interpretations, so this is not a formal accuracy score. Preview shows final-answer excerpts; native outputText, SHA256 and metrics retained. Token rates include reasoning and exclude prefill. These checks do not establish general quality or production readiness.
Sampling audited from every vLLM request: temperature 1, top_p 1, top_k 0, min_p 0, presence/frequency penalties 0, repetition penalty 1, max_tokens 262144; context explicitly 1048576. Two warmups and 5 timed samples, current immutable container, no new tuning. All 7 generated programs, including all 5 timed samples, passed 8 isolated stdlib-only functional cases each. Preview shows final-answer excerpts; native outputText, SHA256 and metrics retained. Token rates include reasoning and exclude prefill. These checks do not establish general quality or production readiness.
Sampling audited from every vLLM request: temperature 1, top_p 1, top_k 0, min_p 0, presence/frequency penalties 0, repetition penalty 1, max_tokens 262144; context explicitly 1048576. Two warmups and 5 timed samples, current immutable container, no new tuning. Arithmetic and four-part completion checked for all 7 answers; the battery question permits pooled (13, $5.80) or equal-share (7, $43.90) interpretations, so this is not a formal accuracy score. Preview shows final-answer excerpts; native outputText, SHA256 and metrics retained. Token rates include reasoning and exclude prefill. These checks do not establish general quality or production readiness.
Sampling audited from every vLLM request: temperature 1, top_p 1, top_k 0, min_p 0, presence/frequency penalties 0, repetition penalty 1, max_tokens 262144; context explicitly 1048576. Two warmups and 5 timed samples, current immutable container, no new tuning. All 7 generated programs, including all 5 timed samples, passed 8 isolated stdlib-only functional cases each. Preview shows final-answer excerpts; native outputText, SHA256 and metrics retained. Token rates include reasoning and exclude prefill. These checks do not establish general quality or production readiness.
Temperature 1, top_p 1, maximum output 262144 tokens; code-v1, 2 warmups, 5 timed samples. All 7 generated programs, including all 5 timed samples, completed and passed eight isolated stdlib-only functional checks each: malformed input, since filtering, percentiles, timezone offsets, empty input, invalid record schema, missing latency, singleton percentiles. Preview shows the final answer only; full raw output and SHA256 retained, and token rates include reasoning. These checks do not establish general model quality or production readiness. Different protocol from the earlier temperature-0 capped runs, which are flagged for quality investigation; no percentage comparison is claimed.
Temperature 1, top_p 1, maximum output 262144 tokens; code-v1, 2 warmups, 5 timed samples. All 7 generated programs, including all 5 timed samples, completed and passed eight isolated stdlib-only functional checks each: malformed input, since filtering, percentiles, timezone offsets, empty input, invalid record schema, missing latency, singleton percentiles. Preview shows the final answer only; full raw output and SHA256 retained, and token rates include reasoning. These checks do not establish general model quality or production readiness. Different protocol from the earlier temperature-0 capped runs, which are flagged for quality investigation; no percentage comparison is claimed.
QUALITY INVESTIGATION: this temperature-0, 8192-token canonical run includes unfinished reasoning; the 8-GPU preview contains a pathological repeated-sentence loop. These token rates are not validated useful-answer throughput or evidence of production readiness. Matched current/reference, DSpark/autoregressive and sampling controls are running; do not use this result to infer model quality.
QUALITY INVESTIGATION: this temperature-0, 8192-token canonical run includes unfinished reasoning; the 8-GPU preview contains a pathological repeated-sentence loop. These token rates are not validated useful-answer throughput or evidence of production readiness. Matched current/reference, DSpark/autoregressive and sampling controls are running; do not use this result to infer model quality.
