💎 Want higher quality? This is the XS (smallest-footprint) body. If you have the VRAM — it's only slightly larger — the full Multimodal-NVFP4-MTP body runs at parity speed with materially higher quality-eval scores, and is the recommended DGX Spark daily-driver. Stay on XS when footprint / VRAM is tight.
The GitHub repo is the source of truth for the production deployment guide, hardware-tuned docker-compose configs, full configuration reference, measured benchmarks, and AGENTS.md — an operator's manual that pre-empts common stale-documentation traps.
🆕 Canonical container — AEON vLLM Ultimate
ghcr.io/aeon-7/aeon-vllm-ultimate:latest (= tag :2026-06-18-v0.23.0-dflashfix; rollback tag :2026-06-11-pr41703) — vLLM 0.23.0 built from source for sm_121a + DFlash high-concurrency fix (drafter block-table slice; scales cleanly to c=64) + PR #40898 (DFlash sliding-window attention) + PR #41703 (prefix-caching corruption-immunity with DFlash) + PR #44389 Triton NVFP4 KV cache (~3× KV capacity) + DFlash + TurboQuant K8V4 + AEON sm_121a patches. This is the single image for all Qwen3.6-27B AEON-Ultimate repos.
The image ENTRYPOINT is /bin/bash, so a docker run must use --entrypoint vllm and then serve … — do not write IMAGE vllm serve (that runs bash vllm serve and fails).
The XS body — even though it ships with native MTP heads — works great with an external DFlash drafter. Pair with z-lab/Qwen3.6-27B-DFlash for big throughput gains over the native MTP path.
🚀 Quickstart (DGX Spark / GB10 — DFlash)
Complete copy-paste recipe: pull the container, pull this model, pull the DFlash drafter, then serve. (Fuller deployment options — dedicated-VRAM Blackwell MTP, env vars, compose — are in the Usage section below.)
bash
1# 1) Pull the canonical AEON vLLM Ultimate container2docker pull ghcr.io/aeon-7/aeon-vllm-ultimate:latest
34# 2) Pull THIS model (fresh)5huggingface-cli download AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS --local-dir ./aeon-model
67# 3) Pull the DFlash drafter (fresh)8huggingface-cli download z-lab/Qwen3.6-27B-DFlash --local-dir ./aeon-drafter
910# 4) Serve (ENTRYPOINT is /bin/bash, so pass --entrypoint vllm then serve …)11docker run --rm --gpus all \12 -v ./aeon-model:/model:ro \13 -v ./aeon-drafter:/drafter:ro \14 --entrypoint vllm ghcr.io/aeon-7/aeon-vllm-ultimate:latest \15 serve /model \16 --quantization modelopt \17 --mamba-cache-dtype float32 \18 --reasoning-parser qwen3 \19 --tool-call-parser qwen3_coder \20 --enable-auto-tool-choice \21 --limit-mm-per-prompt '{"image":4,"video":2}'\22 --mm-encoder-tp-mode data \23 --gpu-memory-utilization 0.65\24 --max-num-seqs 64\25 --max-num-batched-tokens 16384\26 --enable-chunked-prefill \27 --enable-prefix-caching \28 --trust-remote-code \29 --speculative-config '{"method":"dflash","model":"/drafter","num_speculative_tokens":10}'
Lower --gpu-memory-utilization (e.g. 0.69) if the host co-runs other services; on DGX Spark unified memory keep it at 0.6-0.7 — above ~0.8 the shared CPU+GPU pool page-thrashes and stalls the box (even 0.85 stalls). For dedicated-VRAM Blackwell (MTP via the grafted head, no external drafter) see Usage.
📈 Why this image matters for long-context drafting
The z-lab Qwen3.6-27B DFlash drafter is a sliding-window model — 4 of its 5 layers use sliding-window attention (window 2048). aeon-vllm-ultimate:latest (PR #40898) runs those layers as proper SWA; earlier images ran them as full attention, so drafting collapsed once context grew past ~2048 tokens. PR #41703 additionally makes --enable-prefix-cachingcorruption-immune with DFlash. Net: long-context drafting holds up; short-context (<2048, one window) is unchanged.
🏆 Live production bench — DFlash n=10 on aeon-vllm-ultimate:latest
Measured on DGX Spark GB10, aeon-vllm-ultimate:latest, DFlash num_speculative_tokens=10. Lead with acceptance (stable across samples), not single-sample tok/s.
Long-context (~9k-token) draft acceptance — this is the headline win:
Image
~9k-token draft acceptance
pre-fix image (full-attn drafter)
19.7 %
aeon-vllm-ultimate:latest(SWA drafter)
45.0 %(2.3×)
Short-context c=1 acceptance by category (new image, n=10, approximate):
Category
accept
Math
~50 %
Reasoning
~50 %
Extraction
~40 %
Coding
~38 %
Natural
~25 %
Prose
~18 %
Short-context throughput is statistically unchanged vs the prior image — the drafter's sliding window only engages past 2048 tokens, so the win is specifically long-context. (Caveat: single / 3-round samples; short-context rankings are within noise. Acceptance is the stable signal — single-sample tok/s is not.)
🙏 Reference recipe credit: The conv1d-preserved NVFP4 + MTP graft pipeline used to build this XS variant is based on sakamakismile's validated Qwen3.6-27B-NVFP4-MTP series (22K+ downloads). They worked out the modelopt config — including the strategic decision to quantize the GDN projection matmuls to NVFP4 while preserving linear_attn.conv1d at BF16 — and the MTP-head graft technique. We adapted the recipe to AEON-Ultimate's abliterated weights and ship both the conv1d-preserved-only XS variant (matching their footprint) and a heavier regular-MTP variant that additionally keeps the projections at BF16. Full credit for the underlying recipe → sakamakismile.
Fastest 27B export. NVFP4 MTP-XS body + an external DFlash@10 drafter on
ghcr.io/aeon-7/aeon-vllm-ultimate:latest (vLLM 0.23.0) delivers ~42.6 tok/s single-stream
and ~340 tok/s aggregate at c=64, with DFlash draft acceptance ~35 % at short context that
holds ~45 % at long (~9k) context. This is the body the canonical 27B card was benchmarked on.
Measured on DGX Spark GB10 (sm_121a, unified memory), aeon-vllm-ultimate:latest, NVFP4 body served with an external DFlash drafter at num_speculative_tokens=10. The grafted MTP head ships in this repo but sits unused on the Spark — DFlash wins on unified memory (see hardware routing below).
Qwen3.6-27B AEON-Ultimate variant comparison: MTP-XS NVFP4+DFlash vs MTP vs BF16 baseline, single-stream and c=64 aggregate throughput on DGX Spark
MTP-XS is the smallest NVFP4 export and the fastest single-stream of the 27B family, at roughly half the memory of the BF16 baseline.
Per-category single-stream (c=1)
Category
Decode (tok/s)
TTFT (ms)
TPOT (ms)
Prefill (tok/s)
DFlash accept
Coding
42.6
141
23.5
318
34.5 %
Math
55.9
248
17.9
246
48.0 %
Reasoning
49.3
232
20.3
211
41.7 %
Prose
31.2
229
32.1
166
23.2 %
Natural language
34.8
228
28.7
175
26.6 %
Extraction / JSON
57.4
234
17.4
231
49.3 %
Decode speed tracks DFlash acceptance: structured workloads (Extraction, Math, Reasoning) draft well (≈42–49 % accept → 49–57 tok/s); free-form prose drafts less predictably (≈23 % → 31 tok/s). The headline ~42.6 tok/s is the Coding-category single-stream figure.
Aggregate throughput by concurrency
Aggregate throughput vs concurrency (c=1 to c=64) per prompt category for Qwen3.6-27B Multimodal-NVFP4-MTP-XS with DFlash@10 on DGX Spark
Throughput scales cleanly to c=64 (the DFlash high-concurrency fix below removed the prior c≥32 crash). Aggregate peaks at c=64, topping out around ~340 tok/s (Reasoning category); every category climbs monotonically from c=1 → c=64:
Category
c=1
c=8
c=16
c=32
c=64
Coding
42
185
249
262
277
Math
53
221
285
294
303
Reasoning
47
241
301
319
340
Prose
30
130
165
179
193
Natural language
34
148
217
208
230
Extraction / JSON
55
201
250
270
299
Long-context DFlash acceptance
The z-lab DFlash drafter is a sliding-window model (4 of 5 layers use SWA, window 2048). On this image (PR #40898) those layers run as proper SWA, so draft acceptance holds as context grows instead of collapsing past 2k tokens:
Context
DFlash draft acceptance
short (c=1, blended)
~35 %
long (~9k tokens)
45.0 %
This is the headline long-context win — acceptance at ~9k tokens is higher than the blended short-context average. (Pre-fix full-attention image collapsed to ~19.7 % at the same context.)
What we fixed for the DGX Spark
All AEON models run on one unified container — ghcr.io/aeon-7/aeon-vllm-ultimate:latest (vLLM 0.23.0 built from source for sm_121a, merged with the AEON speculative-decoding stack). Two fixes matter most for this card:
Unified container. A single sm_121a image (vLLM 0.23.0) replaces the per-model image sprawl — the same build serves every Qwen3.6-27B AEON-Ultimate variant, with the SM120-family CUTLASS NVFP4/FP8 kernels GB10 actually dispatches to.
DFlash high-concurrency fix. The speculative drafter previously crashed at ≥32 concurrent requests (a padded-vs-unpadded KV block-table shape mismatch in FlashAttention). The fix slices the drafter's block-table to the unpadded batch (block_table[:num_reqs]) — a port of upstream PR #43982, which fixed this for MTP but never for DFlash. The c=32 / c=64 columns above are only measurable because of it.
Full optimization writeup (NVFP4 KV cache, DFlash SWA, sm_121a build, unified-memory tuning): see the container repo.
Stock baseline pending. These figures are on the optimized aeon-vllm-ultimate:latest. There is no stock / vanilla-vLLM baseline for this export yet — a fresh fully-vanilla re-bench (default settings, no speculative decoding, no sm_121a optimizations) is pending and will be added when it completes.
What "XS" means — and what it's not
This is the extra-small footprint sibling of -Multimodal-NVFP4-MTP. XS is not "everything to FP4." It is a deliberate, principled split: the heavy GDN matmul projections drop to NVFP4 (where they're bandwidth-bound and FP4 wins big), while the SSM-critical linear_attn.conv1d kernel stays BF16 (where FP4 has documented stability problems on long-context recurrence).
linear_attn SSM state vectors (A_log, dt_bias, norm.weight)
preserved BF16
preserved BF16 ✅
mtp.* head (grafted bf16 from base, bit-exact verified)
yes
yes
Vision tower
preserved BF16
preserved BF16
Total disk
~27 GB
~21 GB
VRAM footprint at runtime
~28 GB
~22 GB
This is a smart, strategic quantization — not a precision compromise. The conv1d preservation matters: the GatedDeltaNet recurrence depends on the 1D convolution behaving numerically like its training distribution, and FP4 quantization of conv1d has been observed to cause drift on long-context inference in community testing. By keeping conv1d BF16 while quantizing the projections (which are bandwidth-limited matmuls where FP4 is a clean win), we get the ~6 GB footprint reduction without sacrificing the part of the model that's actually fragile under quantization. This is the same principle modelopt's NVFP4_DEFAULT_CFG applies by default and the same recipe sakamakismile validated across his Qwen3.6-NVFP4-MTP series (22K+ downloads).
When to pick which:
Pick the regular variant if you have ≥48 GB VRAM. Even the projection weights at BF16 give a small additional safety margin on long-context recurrence stability.
Pick this XS variant if you have 24–32 GB VRAM (RTX 5090, single GPUs without headroom for full BF16 GDN). The conv1d preservation guarantees the SSM recurrence stays numerically stable; the ~6 GB savings buy meaningful KV-cache headroom on tight GPUs.
We ship both because we have the headroom on RTX PRO 6000 / B100/B200 to run the larger, more numerically-conservative version, and several users on tighter cards have asked for the smaller one. Neither variant quantizes linear_attn.conv1d — that would be a different (and not-recommended) variant we have explicitly chosen not to ship.
The modelopt-format NVFP4 + MTP variant, multimodal-preserved, with linear_attn projections fully quantized, of AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16 — the lossless abliteration of Qwen 3.6 27B (KL 0.000492 vs base, 0/100 refusals, multimodal preserved, hybrid GDN-aware quantization).
Specifically:
Body quantized to NVFP4 via nvidia-modelopt 0.43.0 with NVFP4_DEFAULT_CFG. modelopt format, served by vLLM through --quantization modelopt.
Linear-attn / GatedDeltaNet projections quantized to NVFP4 (this is the XS difference). Only linear_attn.conv1d is kept BF16 (modelopt's default). The community has validated this approach on Qwen3.5/3.6-NVFP4 builds with 22K+ downloads on sakamakismile's reference recipes; we re-ran calibration on our abliterated weights and the model serves correctly.
Vision tower preserved BF16 (333 keys) — correct model.visual.* layout. Multimodal weights load; runtime vision inference validation on this image is pending a GPU window.
MTP head grafted from the baseQwen/Qwen3.6-27B checkpoint (15 tensors, BF16, bit-exact verified). Powers --speculative-config '{"method":"qwen3_5_mtp",...}' for self-speculative decoding without a separate drafter.
Why MTP
Multi-Token Prediction (MTP) lets the model predict multiple future tokens per forward pass via the trained mtp.* head, enabling speculative decoding without a separate drafter model. The acceptance rate is high because the drafter is the model itself — same architecture, same weights, same distribution.
Indicative published numbers (sakamakismile's reference recipe on RTX 5090):
Mean acceptance length: ~3.0-4.0 (compared to DFlash chains of ~2.0-2.3)
Validated benchmarks of the AEON-Ultimate XS variant land in the GitHub repo once measured.
🎯 When to pick this variant — measured hardware routing
The right speculative-decode method depends on memory architecture:
Hardware tier
Recommended variant
Why
DGX Spark / GB10(sm_121a, unified memory)
Either: -NVFP4 (DFlash)(simpler, validated)or this XS body served with --speculative-config '{"method":"dflash",...}'(highest measured throughput — see the acceptance bench above)
Spark prefers DFlash regardless of body. On aeon-vllm-ultimate:latest with DFlash n=10, long-context (~9k) draft acceptance reaches 45.0 % (2.3× the pre-fix image) — see the live bench above. The grafted MTP head in this repo is unused in that path. Never use --speculative-config '{"method":"qwen3_5_mtp",...}' on Spark — MTP loses badly to DFlash on unified memory.
RTX PRO 6000 Blackwell(96 GB dedicated VRAM)
Multimodal-NVFP4-MTP — GDN BF16 for best long-context fidelity, orthis XS variant for ~10 % faster decode
XS measured 111.4 tok/s median vs regular's 101.5 on RTX PRO 6000. Both win against DFlash on dedicated VRAM.
num_speculative_tokens=3 is the canonical setting for qwen3_5_mtp. Higher values diverge the drafter further from the target distribution and acceptance falls.
vLLM serve — DGX Spark (DFlash spec, not MTP — current production recipe)
For DGX Spark, swap the spec method to DFlash. DFlash's block-diffusion drafter is decisively better than MTP's n=3 on unified memory. This is the exact recipe running in production, on the AEON vLLM Ultimate image ghcr.io/aeon-7/aeon-vllm-ultimate:latest (vLLM 0.23.0 + DFlash high-concurrency fix + PR #40898 + PR #41703 + PR #44389).
The image's ENTRYPOINT is /bin/bash, so when launching via docker run you must pass --entrypoint vllm and then serve … (writing IMAGE vllm serve runs bash vllm serve and fails). The vllm serve … arguments are identical either way:
Critical DFlash config rules (learned the hard way):
Use the DEFAULT drafter attention backend — do not add an attention_backend to the spec-config. The default works for Qwen3.6 DFlash on this image (unlike Gemma's DFlash, which needed an explicit flash_attn backend). Leave it out.
Do NOT set --kv-cache-dtype. DFlash's drafter is non-causal (block diffusion) and no vLLM backend supports non-causal + fp8 KV, so KV must stay at default BF16. Forcing fp8 KV will fail to boot.
num_speculative_tokens=10 is the validated default. An n=8/10/12/15 sweep found n=10 best for the Spark voice/chat workload — top aggregate throughput + DFlash acceptance at parity single-stream; the old 12-token setting wins only at very long context (z-lab's published default is 15).
--gpu-memory-utilization 0.69 because this host co-runs Qwen3-ASR (:8001) + Qwen3-TTS (:8002). Keep it ≤ 0.7 when co-hosting; even with vLLM running alone stay at 0.6-0.7 on the DGX Spark's unified memory — above ~0.8 the shared CPU+GPU pool page-thrashes and stalls the box (even 0.85 stalls). BF16 KV is 2× fp8, but full 256k context still fits — KV cache holds 487k tokens / 1.86× concurrency at 262,144 ctx.
Why this recipe needs aeon-vllm-ultimate:latest: the z-lab DFlash drafter is a sliding-window model — 4 of its 5 layers use sliding-window attention (window 2048). This image (PR #40898) runs those layers as proper SWA; earlier images ran them as full attention, so drafting collapsed once context grew past ~2048 tokens. PR #41703 additionally makes --enable-prefix-caching corruption-immune with DFlash. Net: long-context drafting holds up; short-context (<2048, one window) is unchanged. See the live acceptance bench at the top (45.0 % @ ~9k vs 19.7 % pre-fix = 2.3×).
Configuration notes
--quantization modelopt is required for this body (not compressed-tensors — different format).
--speculative-config '{"method":"qwen3_5_mtp", ...}' uses the grafted MTP head; correct for dedicated-VRAM Blackwell. Don't use this on DGX Spark.
--speculative-config '{"method":"dflash", ...}' uses an external DFlash drafter; correct for DGX Spark. The grafted MTP head in this repo sits unused in this path (~0.85 GB dead weight). Don't use this on RTX PRO 6000 or B100/B200 — they prefer MTP.
--gpu-memory-utilization 0.94 is the validated cap on RTX PRO 6000; on DGX Spark keep it at 0.6-0.7 (unified memory page-thrashes above ~0.8 — even 0.85 stalls).
Quantization recipe
Tool: nvidia-modelopt 0.43.0 with NVFP4_DEFAULT_CFG
*linear_attn.conv1d*, *mixer.conv1d*(NVFP4_DEFAULT_CFG default — kept BF16 because FP4 quantization of the SSM 1D convolution causes drift on long-context recurrence; this is the recurrence-critical kernel of the GatedDeltaNet block. Both regular and XS variants preserve this.)
*linear_attn* is NOT broadly excluded (XS difference — the projection matmuls in_proj_qkv, in_proj_z, in_proj_a/b, out_proj get NVFP4-quantized; saves ~8 GB; FP4 is a clean win on bandwidth-bound matmuls)
*visual* (vision tower preservation)
*mtp* (MTP head preservation)
*output_layer*, output.*
MTP graft: 15 tensors copied bf16 from Qwen/Qwen3.6-27B after modelopt export
Pipeline: lna-lab/GGUF-to-NVFP4-SM120 reference recipe, adapted for AEON-Ultimate-BF16 input + separate MTP source
Quantization: NVIDIA TensorRT Model Optimizer (nvidia-modelopt 0.43.0)
Base: Alibaba Qwen team — Qwen/Qwen3.6-27B
License + responsibility
Apache 2.0, inherited from Qwen/Qwen3.6-27B. This is an uncensored model. Read the full User Responsibility & Arbitration Clause on the BF16 source card before deploying. Summary: you implement downstream safety layers (input validation, output filtering, content moderation, audit logging, rate limiting, access controls, human-in-the-loop for high-risk workflows). The model has no opinions of its own — you supply the opinions, the judgment, and the ethics.
☕ Support the work
If this release has been useful, tips are deeply appreciated — they go directly toward more compute, more models, and more open releases.
