Qwen3.8 Flash.
V3. Measured end to end.

Same CIRU weights, updated runtime and serving profile. At 64K, v3 reduces whole-request time by 23.93%, raises prompt throughput by 29.99%, and raises generation throughput by 81.66% versus the previous CIRU runner. At 4K, whole-request time falls 11.80%.

All three arms ran sequentially on Ciru: Ryzen AI Max+ 395 / gfx1151, 128 GB unified memory, NixOS. Identical input token IDs, cold prompt cache, 128 output tokens, one slot and 262,144 context capacity. Each load has an excluded warmup. Previous CIRU is the locally qualified v2.0.1 runner; it was not separately published as a tag.

Previous CIRU uses MTP6, b2048/u512; v3 uses MTP6, b1024/u1024. Both use F16 target KV and Q8 draft KV. Halo uses its unmodified Vulkan runner, Unsloth UD-Q4_K_XL target, EasiiX Q8 head and screened MTP3 profile. Halo generates faster at 4K; the small 4K whole-request difference is not a robust general v3 win. Three v3 loads, two Halo loads and one previous-CIRU load do not establish confidence intervals. This measures complete serving profiles, not runtime changes in isolation.

The additional Ornith difficulty panel combines 22 short IFEval/GSM8K/HumanEval tasks, six Hermes scenarios repeated twice, eight hard tasks with a shared 63K-token history, and short/long coding health checks. It uses cases selected from earlier disagreements and failures; scores do not estimate general benchmark accuracy.

V3 takes 13.59% less total time than previous CIRU, while Halo finishes 43.89 seconds sooner than v3. Long hard-stage time falls 42.98% versus previous CIRU, but output lengths differ. Short-stage time is effectively unchanged. These are complete workload timings, not equal-output generation speedups.

V3 preserves the previous academic pass counts and improves Hermes native full passes from 7/12 to 11/12. Reviewed end states were 11/12, 12/12 and 12/12. Previous CIRU and Halo each have a memory-case wording artifact in the native grader; original grades are retained. All arms passed the short 10-task and long 8-task coding health checks, base and extended tests.

All four profiles passed 20/20 HumanEval base and extended tests on the separate canonical tasks 0–19 panel. Previous CIRU and v3 MTP6 produced identical output token IDs; v3 reduces their summed request time by 6.39%. This is a bounded nonthinking regression test, not a full 164-task result.

MTP6 retains high-acceptance coding speed; MTP2 lowers it to 39.63 tok/s. On the lower-acceptance fixed-output fixture, optional MTP2 instead measured 13.62 s at 4K and 180.87 s at 64K (29.39 / 24.88 generation tok/s). It remains an option, not the general default. The target model verifies the full vocabulary at either depth.

Recall checks recovered both keys with exact cached replay at approximately 8K and 64K for all three main arms. V3 passed 69 QSA mapping/state/guard cases, 33 ROCm operator reference cases, and 30 allocator tests with 198 assertions. A four-prefix diagnostic matched 15,892,480 F32 logits byte-for-byte; it does not establish universal equivalence or long-context task accuracy.

These are the completed H96, MTP depth-1 results behind the earlier release’s EvalScope scores. They used the same released weights, one request at a time and uncapped natural-EOS generation. They have not been rerun on v3.

Quality stages together: 3h 17m 50s. Including the separate performance test: 3h 20m 46s. This is not the 25-minute v3 mixed panel above. No estimated v3 EvalScope wall time is presented as a measurement.