DEV.co
Open-Source LLM · nvidia

Qwen3.6-27B-NVFP4

Qwen3.6-27B-NVFP4 is a 27-billion-parameter quantized language model from NVIDIA, derived from Alibaba's Qwen3.6-27B. It compresses the original model to 4-bit precision using NVIDIA's Model Optimizer, reducing memory footprint by approximately 2.5x while maintaining near-identical accuracy. The model handles text, image, and video inputs up to 262K context length and is optimized for inference on NVIDIA GPUs via vLLM.

Source: HuggingFace — huggingface.co/nvidia/Qwen3.6-27B-NVFP4
18.2B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
538.7k
Downloads (30d)

Key facts

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

FieldValue
Developernvidia
Parameters18.2B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads538.7k
Likes315
Last updated2026-06-30
Sourcenvidia/Qwen3.6-27B-NVFP4

What Qwen3.6-27B-NVFP4 is

A post-training quantized transformer with hybrid attention (Gated DeltaNet and Gated Attention) that uses NVFP4 (4-bit) precision for weights and activations in linear operators. Designed for vLLM inference on Hopper/Blackwell GPU microarchitectures. Calibrated on CNN-DailyMail and Nemotron-Post-Training-Dataset-v2. Evaluated on nine benchmarks including reasoning (MMLU Pro, GPQA, HLE), multimodal (MMMU Pro), coding (SciCode), and agentic tasks (τ²-Bench Telecom). Performance remains within 0.1–0.5 points of FP8 baseline across most benchmarks.

Quickstart

Run Qwen3.6-27B-NVFP4 locally

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

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="nvidia/Qwen3.6-27B-NVFP4")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

Cost-optimized production inference at scale

4-bit quantization reduces memory and compute requirements by ~2.5x while maintaining accuracy within 0.1–0.5% of FP8 baseline. Ideal for high-throughput deployment where cost per token and GPU utilization are critical.

Long-context RAG and document processing

262K context window enables retrieval-augmented generation with large document sets or multi-turn conversation history without re-ranking or segmentation overhead.

Agentic AI systems and chatbots

Evaluated on τ²-Bench Telecom (95.4% accuracy) and instruction-following benchmarks (IFBench 65.5%), demonstrating reliable tool-use and policy adherence for customer-service automation and multi-step reasoning tasks.

Running & fine-tuning it

ESTIMATE: ~6.5–7.5 GB VRAM for inference (27B params × 4 bits ÷ 8 + KV cache overhead). Requires NVIDIA Hopper or Blackwell GPU. Test hardware noted as GB300. Specific throughput (tokens/sec), batch size, and memory scaling are not provided; verify with vLLM profiling on target hardware.

Not documented. Model card does not mention LoRA, QLoRA, or parameter-efficient tuning compatibility. Quantization is post-training only. Contact NVIDIA or consult Model Optimizer documentation to assess fine-tuning feasibility on quantized weights.

When to avoid it — and what to weigh

  • Extreme latency constraints or CPU-only deployment — Model is optimized for NVIDIA GPU inference via vLLM. No CPU or alternative runtime (TGI, llama.cpp) integration is documented. CPU inference will be prohibitively slow.
  • Task requires unquantized model or full precision training — NVFP4 is a post-training quantization. Fine-tuning details and LoRA feasibility are not documented. Unquantized Qwen3.6-27B may be required if quantization loss is unacceptable for your domain.
  • Toxic content or societal bias mitigation is mandatory — Model card explicitly notes the base model contains toxic language and societal biases from internet crawls. Model may amplify biases and return toxic responses, particularly with adversarial prompts. Domain-specific safety validation required.
  • Non-Linux operating systems or non-Hopper/Blackwell hardware — Preferred OS is Linux; supported microarchitectures are Hopper and Blackwell only. Deployment on Windows, macOS, older NVIDIA GPUs, or AMD hardware is not documented.

License & commercial use

Apache 2.0. Permissive OSI-approved license permitting commercial and non-commercial use, modification, and distribution under Apache terms.

Model card states 'This model is ready for commercial or non-commercial use' and is governed by Apache 2.0. Apache 2.0 is a permissive OSI license that explicitly allows commercial use. No additional restrictions or proprietary licensing found in provided data. However, verify compliance with Alibaba's underlying Qwen3.6-27B license and any NVIDIA ModelOpt terms.

DEV.co evaluation signals

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

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

Model card does not claim security properties. Practical considerations: (1) Quantization may introduce subtle numerical shifts; validate outputs in security-sensitive tasks. (2) vLLM runtime and CUDA stack are attack surface; keep driver and libraries updated. (3) Long context (262K tokens) exposes prompt-injection and context-leaking risks; sanitize user input. (4) Model inherits toxic language and bias from training data; use content filtering for user-facing deployments. No vulnerability disclosure process or security audit results provided.

Alternatives to consider

Qwen3.6-27B-FP8

Official baseline (Alibaba/HuggingFace). Higher precision (8-bit) trades ~2x memory savings for potential 0.1–0.5% accuracy gains on some benchmarks. Better for precision-critical tasks.

Qwen3.6-27B (unquantized)

Full 27B model at 16-bit precision. Larger footprint but allows fine-tuning and inference without quantization artifacts. Choose if model accuracy loss is unacceptable.

Llama 3.1 70B quantized (e.g., via TGI or llama.cpp)

Larger parameter count (70B vs 27B) with broader third-party runtime support (CPU-compatible quantization, multi-platform serving). Trade-off: higher compute cost, not NVIDIA-optimized.

Software development agency

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

Qwen3.6-27B-NVFP4 combines 2.5x memory savings with near-baseline accuracy for high-throughput inference. Verify NVIDIA GPU compatibility, test vLLM integration, and validate quantization performance on your use case before production rollout.

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-NVFP4 FAQ

Can I use this model commercially?
Yes. The model is governed by Apache 2.0, which explicitly permits commercial use. However, verify that you comply with Alibaba's Qwen3.6-27B base license and any NVIDIA ModelOpt terms.
What GPU do I need?
NVIDIA Hopper or Blackwell microarchitecture (H100, H200, B100, etc.) running Linux. Estimated ~6.5–7.5 GB VRAM for inference. Test hardware cited as GB300. Confirm vLLM compatibility and memory footprint on your target GPU before production.
Can I fine-tune this quantized model?
Not documented. The model is post-training quantized; fine-tuning capability and LoRA compatibility are not stated. Contact NVIDIA or consult Model Optimizer docs. You may need to fine-tune the unquantized base model and re-quantize.
What is the context length?
Up to 262K tokens (262,144). Evaluated on AA-LCR (long-context recall) benchmark. Confirm your vLLM --max-model-len configuration matches deployment needs.

Work with a software development agency

DEV.co helps companies turn open-source tools like Qwen3.6-27B-NVFP4 into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Ready to Deploy Quantized LLMs at Scale?

Qwen3.6-27B-NVFP4 combines 2.5x memory savings with near-baseline accuracy for high-throughput inference. Verify NVIDIA GPU compatibility, test vLLM integration, and validate quantization performance on your use case before production rollout.