Skip to content

Qwen3.8-Flash-Next benchmarks

Compare Qwen3.8-Flash-Next inference speed on Olares One, NVIDIA DGX Spark and Mac Studio M5 Max. Review IFBench scores for three quantizations tested on Olares One.

We tested five quantizations with Strata 0.2.10. Coverage varies by device and task.

Inference engine
Strata 0.2.10
Tested quantizations
UD-IQ4_XS, IQ3_XXS, IQ3_S, IQ2_XS, Q2_0
Quality evaluation
IFBench on Olares One

IFBench

IFBench is an instruction-following benchmark from Ai2. It tests whether models follow explicit requirements in their responses.

300 prompts, xhigh reasoning, 65,536-token output limit

Each configuration was evaluated once on Olares One. Scores may vary between runs. IQ3_S and Q2_0 were tested for speed only. Truncated outputs reached the 65,536-token limit.

Evaluation time

Lower is better
MachineConfigurationScoreTruncatedOutput per questionToken generationEvaluation time
OfficialModel card ↗Original weights, model card ↗81.3————

One run varies by about ±2.3 points at this question count, so a score above the official one is noise, not a gain. Evaluation time is about questions × output per question ÷ token generation speed, one request at a time. It also includes loading and scoring.

Inference speed

Prompt processing and single-request token generation across five quantizations tested with Strata.

20 prompts, medium reasoning, 3,072-token output limit

One request at a time. Hardware and Strata settings differ between systems. Q2_0 has speed results only, with no IFBench evaluation.

Token generation, single request

Higher is better
Mac Studio M5 MaxStrata, Q2_0
130 tok/s
DGX SparkStrata, Q2_0
80 tok/s
DGX SparkStrata, IQ3_XXS
68 tok/s
DGX SparkStrata, IQ3_S
59 tok/s
DGX SparkStrata, UD-IQ4_XS
52 tok/s
Mac Studio M5 MaxStrata, IQ3_XXS, no MTP
47 tok/s
Mac Studio M5 MaxStrata, IQ3_S, no MTP
41 tok/s

Prompt processing (prefill)

Client-observed prompt processing speed, median of three runs per prompt length

Calculated as prompt tokens divided by time to first token. This differs from the engine-reported metric on the Qwen3.8-27B page. Use the same metric when comparing models.

Prompt length in tokens. Speed in tokens per second.

Test configurations

All tests used Strata 0.2.10 with engine 0.1.56 and one request at a time. Quantization and runtime settings affect both speed and quality.

  • Olares One used MTP speculative decoding for all five quantizations. Fused prefill was disabled for this GPU. The setting that reserves output tokens for the final answer was also disabled.
  • NVIDIA DGX Spark used MTP and Strata's default launch settings. The setting that reserves output tokens for the final answer was enabled, so output lengths are not directly comparable with Olares One.
  • The tested Mac Studio M5 Max has 64 GB of unified memory. Q2_0 used MTP. IQ3_XXS and IQ3_S streamed experts from SSD without MTP under Strata's default settings.

IQ2_XS failed to load on DGX Spark in this test and is not offered by Strata on macOS. UD-IQ4_XS was not tested on the Mac. These missing results do not represent zero performance.

For comparison, see Qwen3.8-27B benchmarks. Its IFBench results on Olares One average three runs for the NVFP4 LOW configuration. Flash-Next scores here are single runs and do not establish a consistent quality advantage.