DEV.co
Open-Source LLM · AEON-7

Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS

Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS is a 27B parameter multimodal LLM quantized to NVFP4 precision with multi-token prediction (MTP) capability. It is designed for high-throughput inference on NVIDIA hardware (DGX Spark, Blackwell GPUs) with optional DFlash speculative decoding. The model is Apache 2.0 licensed, ungated, and includes deployment recipes via Docker. It shows ~42.6 tok/s single-stream and ~340 tok/s at c=64 concurrency on DGX Spark.

Source: HuggingFace — huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS
17.1B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
37.4k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperAEON-7
Parameters17.1B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads37.4k
Likes53
Last updated2026-07-03
SourceAEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS

What Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS is

27B parameter multimodal model (text + image inputs), quantized via NVIDIA ModelOpt to NVFP4 with conv1d preservation in linear attention layers. Ships with grafted MTP heads for speculative decoding on dedicated-VRAM Blackwell; recommended to pair with external DFlash drafter on unified-memory systems (DGX Spark). Served via canonical vLLM 0.23.0 container (ghcr.io/aeon-7/aeon-vllm-ultimate:latest) with high-concurrency DFlash fix (c=64 validated), prefix-caching, and sliding-window attention support for long-context drafting. Context length: Unknown. Baseline model: Qwen 3.6. Tags indicate: abliterated (safety guardrails removed), multi-token-prediction, speculative-decoding, hybrid-attention (mamba + gated-deltanet), conversational, English-focused.

Quickstart

Run Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

High-throughput API serving

Quantized NVFP4 body scales to c=64 concurrency with aggregate ~340 tok/s on DGX Spark. Ideal for batch inference, multi-tenant SaaS, or internal API endpoints where throughput is the primary constraint.

Structured extraction and JSON generation

57.4 tok/s single-stream decode with 49.3% DFlash acceptance in extraction/JSON tasks. Well-suited for document parsing, data extraction pipelines, and LLM-as-judge workloads.

Reasoning and code generation with long-context drafting

Math and reasoning categories show 48–50% DFlash acceptance at short context, holding ~45% at ~9k-token context with SWA drafter. Suitable for multi-turn reasoning chains, code review, and iterative problem-solving where draft acceptance remains stable.

Running & fine-tuning it

Estimated VRAM: 16–20 GB (NVFP4 quantization at 27B parameters; card states 'roughly half the memory of BF16 baseline'). Optimized for NVIDIA sm_121a (Blackwell, DGX Spark), sm_120, sm_100; supports RTX 5090, RTX Pro 6000, B100, B200. Minimum 32 GB host VRAM recommended for unified-memory systems; GPU memory utilization up to 0.85–0.88 per deployment guidance. Requires NVIDIA CUDA-capable hardware and Docker for canonical serving.

Not explicitly addressed in the card. Model is a graft of quantized NVFP4 body with MTP heads; fine-tuning viability (LoRA/QLoRA compatibility, merge stability) is Unknown. Abliterated weights may complicate alignment if re-training for safety. Recommend consulting AEON-7 GitHub (referenced in card) or contacting developer before committing to fine-tuning workflows.

When to avoid it — and what to weigh

  • Censorship-sensitive applications requiring safety guarantees — Model is explicitly abliterated (guardrails removed). Not suitable for production environments where content policy enforcement is mandatory (e.g., consumer-facing chat, moderated platforms).
  • Latency-critical single-request scenarios — TTFT ranges 141–248 ms depending on category; decode latency (TPOT 17.9–32.1 ms) is acceptable for batch but may be marginal for strict <100ms RTL targets. Prose category peaks at 32.1 ms per token.
  • Resource-constrained environments without enterprise GPU — Requires modern NVIDIA hardware (DGX Spark, Blackwell, RTX 5090, RTX Pro 6000, B100/B200). VRAM footprint estimated ~16–24 GB (NVFP4); no official mobile, CPU, or legacy GPU guidance provided.
  • Multilingual or non-English workloads — Tags and deployment focus are English-only (en tag). Capabilities in other languages are Unknown.

License & commercial use

Apache 2.0 (apache-2.0). Apache 2.0 is a permissive OSI-approved license permitting commercial use, modification, and distribution under stated terms (patent grant, liability disclaimer, license inclusion). No additional restrictions stated on model card.

Apache 2.0 license permits commercial use, provided licensee includes a copy of the license and states material changes. However, model is explicitly abliterated (safety guardrails removed), which may conflict with internal risk/compliance policies in regulated industries (e.g., finance, healthcare, government). Commercial buyers must: (1) review Apache 2.0 terms, (2) independently assess abliterated behavior against their compliance posture, (3) decide whether removal of safety layers is acceptable for their use case. No warranty or SLA provided; production deployment at buyer's risk. Recommend legal review before enterprise deployment.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Model is abliterated (safety guardrails intentionally removed). Potential for: (1) harmful content generation (biased, abusive, illegal guidance) if prompted adversarially, (2) injection attacks if deployed in multi-tenant or user-facing settings without prompt guardrails. No built-in filtering. Mitigation: layer external content policy (LLM-as-judge, regex filters, prompt instrumentation) before production deployment. NVFP4 quantization introduces minimal novel attack surface vs. FP32/BF16 (quantization artifacts, not security holes). No known CVEs or audit reports cited. Recommend threat modeling and red-teaming before production launch, especially in regulated contexts.

Alternatives to consider

Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP (full variant)

Same base model, heavier (keeps projections at BF16), claimed 'materially higher quality-eval scores' and 'parity speed' on DGX Spark if VRAM permits. Trade: +2–4 GB VRAM for potential accuracy gain.

Llama 3.1 70B or Qwen 3.6 70B (base, FP8/GGUF quantized)

Larger, uncensored alternatives with broader community support and LoRA fine-tuning recipes. Slower on same hardware but may offer better long-context stability and multilingual coverage; trade higher latency for generality.

z-lab Qwen3.6-27B-DFlash (drafter-only)

Lightweight (2.5B) speculative drafter; pairs with this model for external drafting. Not a standalone replacement but a complementary component to boost long-context throughput if unified-memory latency is critical.

Software development agency

Ship Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS with senior software developers

This model is production-ready on modern NVIDIA hardware with containerized vLLM serving. If you need extraction, code generation, or high-concurrency batch inference with quantized efficiency, pull the Devco guide on private LLM deployment or contact our AI engineering team for hardware sizing and serving architecture review.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS FAQ

Can I use this model commercially?
Apache 2.0 license permits commercial use (copy, modify, distribute) provided you include the license and disclose material changes. However, the model is abliterated (safety guardrails removed), which may violate internal policies or regulatory requirements in finance, healthcare, or government. Conduct legal/compliance review before production deployment. No warranty or SLA is provided by the developer.
What is the estimated memory footprint?
Estimated ~16–20 GB VRAM in NVFP4 precision (card states 'roughly half the memory of BF16 baseline'). Exact size depends on batch size and max sequence length. Deployment guidance recommends 0.85 GPU memory utilization on DGX Spark; do not exceed 0.88. For co-hosted workloads, use 0.69 or lower.
Should I use the XS variant or the full MTP variant?
Use XS if VRAM is tight (~16–18 GB). Use the full Multimodal-NVFP4-MTP variant (~20–24 GB estimated) if you have capacity; it ships higher quality-eval scores at parity speed on DGX Spark. Both ship MTP heads but perform better with external DFlash drafter on unified-memory hardware.
What context length does this model support?
Context length is not stated in the model card. Recommend checking the GitHub deployment guide or contacting AEON-7. Long-context drafting (up to ~9k tokens tested) is supported via external DFlash drafter with sliding-window attention.

Work with a software development agency

Adopting Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-NVFP4-MTP-XS is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.

Ready to Deploy High-Throughput Inference?

This model is production-ready on modern NVIDIA hardware with containerized vLLM serving. If you need extraction, code generation, or high-concurrency batch inference with quantized efficiency, pull the Devco guide on private LLM deployment or contact our AI engineering team for hardware sizing and serving architecture review.