BetterBench 0.6.0 · 14 Oct 2026 · 08:37 · kivu-chizungu

qwen3.8-flash-next-iq3_s

qwen3.8-flash-next-iq3_shttp://192.168.178.230:8041/v1corpus v1.05 passes/cattemp 0.7cold prefix cache (nonce)engine: stratagpu: rx9070cpu: ryzen5-3600ram: ddr4-2400quant: IQ3_Sctx: 131072strata: 0.1.43.1+chat-structured-toolsreasoning_effort: lowSTRATA_HIP_WMMA: 1
Combined decode 50.6 t/s weighted across categories
Combined update p99 80.5 ms the stutter between stream updates
Combined TTFT p50 2,264 ms single-stream, batch = 1
Aggregate @ 8 concurrent 38.5 t/s 48/48 ok
Prefill @ 64,000 tok 1,598 t/s median prompt processing
Decode throughput by category Median per-pass decode t/s at batch = 1, 5 passes per category. The dashed line is the weighted combined score.

Hover a bar for its IQR and coefficient of variation — a high CV means the category's passes disagree, so read small differences there with care.

Stream-update gap by category This server packs several tokens into one stream update, so there is no per-token latency to plot. These are the measured wall-clock gaps between updates: p50 is the typical rhythm, p99 the stutter. Lower is better.
update p50update p99

Hover a bar for the tokens landing per update. p99 is an upper bound on any pause a reader feels; it is not comparable across servers that pack different numbers of tokens into an update.

Aggregate throughput under concurrency Total tokens/sec across all in-flight requests as load rises — where this flattens is the throughput knee.
Time-to-first-token under concurrency Queueing shows up here first: p50 is the typical wait, p99 the tail.
TTFT p50TTFT p99

With a modest request count per level, p99 rests on very few observations — treat a spike as a prompt to re-run, not a conclusion.

Prompt processing throughput by depth Median prefill t/s (prompt tokens ÷ TTFT) at increasing input depth, cold prefix cache. The band spans the 1% low to the 99% high.
Full numbers
Single-stream, batch = 1 · several tokens per stream update
categorypassesTTFT p50TTFT p99update p50 (ms)update p99 (ms)tok/updatedecode med±IQRCV
chat52,289.32,649.2†1.283.31.0446.26.011.0%
code52,310.52,347.4†1.078.71.0452.90.61.6%
file edit52,559.12,877.6†0.896.51.0655.52.94.5%
json52,382.12,436.1†0.986.61.0451.45.29.0%
math52,294.92,512.3†0.974.21.0259.93.46.3%
prose51,996.52,280.9†22.269.11.0148.70.91.2%
reasoning52,224.12,398.6†1.081.51.0244.88.517.5%
summarization52,132.52,552.0†0.976.21.0351.85.011.9%
Concurrency sweep
levelok/reqaggregate t/sTTFT p50TTFT p99per-req decode med
148/4839.22,322.32,833.0†50.9
248/4839.411,165.124,721.7†51.1
448/4838.730,729.973,108.2†49.2
848/4838.550,831.5268,746.9†50.2
Reasoning / answer split · TTFA is time to the first answer token
categoryruns w/ splitreasoning share (est)TTFA p50 (ms)never reached answer
chat5/571%—3/5
code5/556%—1/5
file edit5/562%4,814.50/5
json5/559%—1/5
math5/565%—2/5
prose5/545%—1/5
reasoning5/564%—1/5
summarization5/555%4,194.50/5
Prefill sweep · cold prefix cache
target depthprompt tok medTTFT p50PP 1% lowPP medianPP 99% high
2,0001,5443,781.9301.1†407.4441.6†
8,0005,9224,736.71,231.3†1,252.31,259.0†
16,00011,8308,351.51,408.8†1,415.51,752.8†
32,00023,57816,876.01,389.2†1,396.71,400.6†
64,00047,04429,440.01,595.1†1,598.21,690.1†