Qwen3.6-27B AEON Uncensored · vLLM-ultimate (custom) · NVFP4 + DFlash · conc 32

status: done

Configuration

ModelAEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored
CompanyAlibaba
FamilyQwen
Parameters27B (dense) + DFlash external drafter
EnginevLLM (aeon-vllm-ultimate custom container) + DFlash (z-lab drafter, num_speculative_tokens 12) (speculative decoding)
Quant / precisionNVFP4 (XS mixed-precision)
Why this quantSame AEON NVFP4-XS model served via the card's "DGX Spark production" recipe — custom container ghcr.io/aeon-7/aeon-vllm-ultimate:latest (vLLM 0.23.0) + external z-lab/Qwen3.6-27B-DFlash drafter, DFlash num_speculative_tokens=12. User's explicit request. SAFETY — untrusted image; NO creds, models READ-ONLY.
DownloadAEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s180.99
Decode tok/s184.21
Peak memory (GB)111.34 (system MemAvailable delta (10s sampling) — custom vLLM static KV reservation (util 0.85))
Completed2026-06-24 01:35 +0800

Full run command

# UNTRUSTED third-party container — NO creds, full cached repo root + drafter mounted READ-ONLY.
scripts/bench-aeon-ultimate-serving.sh 65536 32 1000 900 256
# 668 prompts completed before the 900 s cap (0 errors). TTFT median 41115.7 ms.

qwen3.6-27b-aeon-uncensored Alibaba Qwen NVFP4 16-40B conc-32

Notes

Custom AEON-ultimate container + DFlash, conc 32. Part of the 1/8/32 sweep on the card’s “DGX Spark production” recipe (custom vLLM 0.23.0 + external z-lab DFlash drafter). Direct A/B vs the native-MTP-on-stock-vLLM result (303 tok/s @ conc-32) — does the custom container + DFlash beat stock + MTP at the same batch? The card claims ~340 tok/s @ c=64 / ~45% DFlash accept. See the conc-1 page for the safety posture (untrusted image, no creds, models read-only) and drafter details.

  • Result (conc 32): prefill 180.99 tok/s, decode 184.21 tok/s aggregate; 668 prompts completed before the 900 s cap, with 0 errors. Median TTFT 41115.7 ms. This is the best aggregate throughput of the custom AEON sweep, but it still falls well short of both the benchmark completion target and the card’s own production claims.
  • Acceptance: DFlash remained in roughly the same regime as the lower-concurrency points: mean acceptance length ~2.9-4.4 and avg draft acceptance ~16-28%. The first draft positions accept reasonably, later positions decay quickly. So the custom drafter is doing work, but it is not achieving the kind of acceptance profile that would justify a 12-token speculative budget.
  • Bottom line vs the stock path: this custom container + external DFlash run does not beat the already-recorded native-MTP-on-stock-vLLM path at the same concurrency. It stays cap-bound and under-delivers on throughput even at the most favorable point of its own 1/8/32 sweep.