I’m sharing Juud_engine, a free MIT-licensed experimental fork of Strata v0.1.39 for local Qwen3.8-Flash-Next inference. It preserves Strata’s local OpenAI-compatible API and adds two opt-in CPU-side decode changes:
I compared a source build with pinned Strata on one RTX 4090 (i7-13700), using the same Qwen3.8-Flash-Next IQ3_S pack. Each request generated 256 tokens. Four prompts covered code contexts of 4K, 32K and 128K tokens, plus a Korean 2K prompt. In five paired runs per prompt, median paired decode throughput was 3.1–9.1% higher; paired full-response latency improved 0.5–5.6%.
Important limits: Normal A/B output text matched in only 5/20 pairs. Different continuations and MTP acceptance can change timing, so those runs do not isolate the effect of the code. In a separate three-pair-per-prompt control, outputs matched in 12/12 pairs and decode throughput was 9.6–15.6% higher. That control used different cache and PCIe settings from the normal run, so I do not combine the results. Both changes were enabled together; their individual contributions are unmeasured. H100 and multi-user performance are also unmeasured.
The repository has the source diff, build instructions, model/binary hashes, request/output hashes, pair-level metrics and a verification script: GitHub - bnkaiteam/Juud_engine: Experimental Strata fork for Qwen3.8-Flash-Next with paired RTX 4090 benchmarks and opt-in CPU decode optimizations · GitHub . A source build and the RUN-JUUD wrapper are needed to enable the tested path; the inherited default setup can fetch a Strata prebuilt instead. Model weights are separate and have their own license.
I’d value independent replications (including negative results), suggestions for a per-change ablation, and feedback on the benchmark method.
Disclosure: I am sharing my project, and this post was drafted with AI assistance. The figures were checked against the published paired data.