Qwen3.6-27B AEON Uncensored · vLLM · NVFP4 + MTP (XS)

# TRUSTED PATH ONLY: native MTP on STOCK vLLM nightly-aarch64 (no custom container). The card's
# "DGX Spark production" recipe uses a CUSTOM third-party container (ghcr.io/aeon-7/aeon-vllm-ultimate)
# + a DFlash external drafter — DECLINED (do not run an untrusted container). The repo's own native-MTP
# form runs on stock vLLM with --quantization modelopt:
scripts/bench-vllm-serving.sh AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS \
  65536 32 1000 900 256 \
  --trust-remote-code --quantization modelopt \
  --speculative-config '{"method":"qwen3_5_mtp","num_speculative_tokens":3}'
# 1000/1000 prompts, 0 errors, 843.9 s (did NOT hit 900 s cap — full coverage). ready after 601 s.
# vLLM maps method qwen3_5_mtp -> mtp; draft arch resolves as Qwen3_5MTP; quantization=modelopt_fp4.
# TTFT median 844 ms, TPOT median 94.9 ms (valid here — no reasoning-parser).

DONE — AEON-7 uncensored Qwen3.6-27B, NVFP4-MTP-XS, on stock vLLM (trusted path only). An individual-uploader abliterated derivative of Qwen3.6-27B, added at the user’s explicit request. Throughput-only benchmark (ShareGPT general chat) — a benign inference-rate measurement; the model’s lack of refusals is irrelevant to tok/s. The native-MTP path serves cleanly on stock vLLM, no custom container needed.

• Result (conc 32): prefill 312.3 / decode 303.32 tok/s aggregate; 1000/1000 prompts, 0 errors, 843.9 s (did not hit the 900 s cap → full coverage). Peak mem 108.58 GB. TTFT median 844 ms, TPOT median 94.9 ms.
• XS packing + grafted MTP head load on stock vLLM — no rejection. vLLM resolves the arch as Qwen3_5ForConditionalGeneration, maps method:qwen3_5_mtp → mtp (deprecation alias), loads the draft as Qwen3_5MTP, and accepts the mixed-precision modelopt_fp4 weights (GDN proj NVFP4 / conv1d+vision BF16). This confirms the AEON XS format is serviceable on the stock engine without the custom path.
• MTP acceptance matches stock: mean accept-len ~3.0 (range 2.7–3.6, median ~3.0), avg draft acceptance ~66%, per-position ~0.84/0.68/0.54 — indistinguishable from stock Qwen3.6-27B MTP (~3.0/~66%). Abliteration is a weight tweak; it does not change MTP draft quality.
• Speed vs stock: decode 303 tok/s lands ~11% above the stock Qwen3.6-27B NVFP4+MTP run (274 tok/s). Not a clean comparison — different (abliterated) weights and a different quant recipe (XS mixed-precision vs the stock unsloth NVFP4 packing). Treat as “serves at least as fast as stock,” not a validated speedup; the distinct model tag keeps it out of the clean stock-27B group.
• DECLINED — custom container + DFlash path. The card’s DGX Spark recipe pulls a third-party image (ghcr.io/aeon-7/aeon-vllm-ultimate:latest) + a /drafter for method:dflash,num_speculative_tokens:12. Untrusted containers are not run here; only the native qwen3_5_mtp path on stock nightly-aarch64 was benchmarked. (The card’s “340 tok/s @ c=64, 45% DFlash accept @9k” come from that custom path and are not reproduced here.)
• Trust & safety: individual uploader (AEON-7, 41k dl/mo, 48 likes), Apache-2.0, modelopt NVFP4 confirmed. Run per the user’s explicit request — same exception as cosmicproc/shieldstar.