DEV.co
Open-Source LLM · sakamakismile

Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP

Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP is a 27-billion-parameter text-generation model optimized for deployment on 4× NVIDIA Blackwell GPUs. It combines hybrid attention (linear + periodic full), built-in multi-token prediction (MTP) for speculative decoding, and NVFP4 W4A4 quantization to fit ~7.2 GiB per GPU. The model supports reasoning tags, XML tool-calling, and long contexts (up to 256K tokens). Apache 2.0 licensed, non-gated, and designed for self-hosted inference via vLLM.

Source: HuggingFace — huggingface.co/sakamakismile/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP
17.1B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
38.6k
Downloads (30d)

Key facts

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

FieldValue
Developersakamakismile
Parameters17.1B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads38.6k
Likes62
Last updated2026-05-31
Sourcesakamakismile/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP

What Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP is

A Qwen3.5-family hybrid-attention model (17.1B parameters) with an integrated MTP draft head for speculative decoding. Quantized via NVIDIA's modelopt (NVFP4, W4A4, group size 16) on weights and core linear layers; embedding, routers, and lm_head remain higher precision. Supports reasoning via `<think>` tags, XML-formatted tool calls, and multimodal token inputs (though served as text-generation). Benchmark: ~81–83 tok/s single-stream (TP=4), scaling to ~880 tok/s at 24 concurrent requests (64K context). KV-cache uses FP8; context window up to 262K tokens. Requires vLLM ≥0.22 with qwen3_5 architecture and MTP support.

Quickstart

Run Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP locally

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

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="sakamakismile/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP")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 reasoning and tool-use workloads

Achieves ~880 tok/s aggregate throughput at 24 concurrent requests and supports built-in reasoning + XML tool parsing. Ideal for agentic systems, multi-turn reasoning, or batch API services where latency per individual request can be moderate (37–40 ms per token).

Long-document and retrieval-augmented generation (RAG)

Hybrid attention keeps KV-cache cost low; 256K context at 16-way concurrency shows near-linear scaling. Suited for document analysis, legal/financial contracts, or RAG pipelines where context length is a constraint.

Self-hosted private LLM deployment

Apache 2.0 license, non-gated, with full Docker/vLLM containerization. Deployment details and performance tuning are well-documented. Fits on 4× RTX PRO 2000 (affordable compared to A100/H100 clusters for this scale).

Running & fine-tuning it

4× NVIDIA RTX PRO 2000 Blackwell (16 GB GDDR6 each), connected via PCIe (NVLink not required). Model weights: ~7.2 GiB per GPU (NVFP4 W4A4). KV-cache: ~2–6 GiB per GPU depending on batch size and context length (FP8). Total GPU memory: 64 GB (4× 16 GB). Host: 32 GB shared memory (shmmax) for vLLM V1. CUDA 13.x + NVIDIA Container Toolkit for containerized deployment. Bare-metal setup also supported with matching toolkit.

Not documented. The checkpoint is NVFP4-quantized (W4A4); unquantized base weights and fine-tuning procedures are not provided. QLoRA on quantized weights is not explicitly supported. Domain adaptation would require either: (1) access to the unquantized Qwen3.5-27B, apply quantization post-training; or (2) evaluate inference-time prompt engineering / in-context learning. Requires review with the developer (sakamakismile) for training-time details.

When to avoid it — and what to weigh

  • Single-GPU or dual-GPU inference — Model requires TP=4 sharding (4× 16 GB GPUs). Cannot run on 2× RTX 6000 or a single H100 without significant refactoring or post-hoc LoRA/pruning.
  • Sub-millisecond latency requirements — Per-token latency at max throughput is 37–40 ms per request; single-stream latency is ~12 ms per token (~83 tok/s). Not suitable for real-time interactive chat where sub-50 ms end-to-end response time is critical.
  • Model fine-tuning or domain adaptation at scale — Quantized W4A4 model is not explicitly documented for QLoRA or LoRA training. Fine-tuning feasibility and convergence are unknown; requires custom integration or access to the unquantized checkpoint.
  • Environments without NVIDIA CUDA 13.x + Blackwell GPU support — NVFP4 kernels and Triton compilation require Blackwell (SM120) or newer. No CPU or AMD GPU support documented. Older NVIDIA architectures (A100, H100) may lack required Triton kernels for NVFP4.

License & commercial use

Apache License 2.0. Permissive, OSI-approved open-source license. Allows commercial use, modification, and distribution under the terms of Apache 2.0.

Apache 2.0 is a permissive OSI license that permits commercial use and derivative works, provided that the Apache 2.0 license text and any NOTICE file are included. No explicit restrictions on commercial deployment, SaaS offerings, or proprietary modifications are stated in the license itself. However, verify upstream dependencies (vLLM, NVIDIA modelopt, PyTorch) for any incompatible licenses if bundling a closed-source service.

DEV.co evaluation signals

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

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

Quantization (NVFP4 W4A4) reduces model transparency but does not eliminate inference-time risks. No explicit security audit, adversarial robustness analysis, or input sanitization recommendations are stated. Deployment on self-hosted infrastructure mitigates data exfiltration risks vs. managed cloud APIs. Reasoning tags (`<think>`) and tool-call XML parsing introduce attack surface if untrusted input is fed without validation. Consider input length limits, prompt injection safeguards, and monitoring of tool-call payloads if exposed to end-users. Standard LLM safety practices (content moderation, rate limiting) apply.

Alternatives to consider

Meta Llama 3.1-70B (quantized, e.g. Ollama/llama.cpp)

Larger model (70B vs. 27B), may offer stronger reasoning and multi-lingual support. Llama 2/3 have broader community and open-source tooling (llama.cpp, Ollama). Requires more GPU memory (2–4× H100/A100); quantization reduces model size but less aggressive than NVFP4.

Qwen2.5-72B-Instruct (unquantized or bfloat16)

Newer Qwen family (2.5 vs. 3.5), larger model size, simpler (no MTP/hybrid attention). Standard bfloat16 inference via vLLM/TGI; broader framework support. Requires ~144 GiB VRAM unquantized or ~36 GiB quantized (1-bit/W4A8).

Deepseek-V3 (MoE, long-context)

Mixture-of-experts architecture allows sparse activation, lower per-token compute. Native long-context (128K+). If latency and throughput at fixed hardware are critical, MoE may offer better scaling. Licensing and quantization status require review.

Software development agency

Ship Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP with senior software developers

Huihui-Qwen3.6-27B-abliterated offers production-grade inference on modest hardware (4× RTX PRO 2000). Start with the provided Docker Compose setup, benchmark on your data, and evaluate reasoning/tool-calling for your agentic or RAG workload. Contact the developer or Devco for deployment support and fine-tuning feasibility reviews.

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.

Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP FAQ

Can I use this model commercially?
Yes. The model is licensed under Apache 2.0, which permits commercial use, distribution, and modification. Ensure you comply with Apache 2.0 attribution (include license text and NOTICE file) and review any upstream dependency licenses (vLLM, NVIDIA modelopt, PyTorch) if bundling into a closed-source product.
What GPU hardware do I need?
4× NVIDIA RTX PRO 2000 Blackwell (16 GB GDDR6 each). The model is sharded across 4 GPUs via tensor parallelism (TP=4) and needs ~7.2 GiB per GPU for weights plus ~2–6 GiB KV-cache. PCIe interconnect is sufficient; NVLink not required. Other NVIDIA GPUs with Blackwell support (e.g., B100) may work if they have ≥16 GB VRAM and Triton NVFP4 kernel support.
How fast is the inference?
Single-stream (1 request): ~81–83 tokens/second. At 24 concurrent requests (64K context): ~880 tokens/second aggregate (~37 tokens/second per request). Per-token latency scales with batch size and context. Speculative decoding via MTP (n=3) improves throughput further by accepting multiple tokens per step; acceptance rates are not explicitly reported.
Can I fine-tune or LoRA this model?
Not documented. The checkpoint is NVFP4-quantized (W4A4); QLoRA on quantized weights is not explicitly supported. Fine-tuning would require either the unquantized base model (Qwen3.5-27B) or confirmation from the developer that the quantized version can be trained. Recommend contacting sakamakismile or the vLLM team for guidance.

Software developers & web developers for hire

Adopting Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP 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 a Quantized Reasoning LLM?

Huihui-Qwen3.6-27B-abliterated offers production-grade inference on modest hardware (4× RTX PRO 2000). Start with the provided Docker Compose setup, benchmark on your data, and evaluate reasoning/tool-calling for your agentic or RAG workload. Contact the developer or Devco for deployment support and fine-tuning feasibility reviews.