DEV.co
Open-Source LLM · nvidia

Qwen3.5-122B-A10B-NVFP4

Qwen3.5-122B-A10B-NVFP4 is NVIDIA's quantized version of Alibaba's 122B mixture-of-experts language model. It reduces model size and GPU memory by ~4x using NVFP4 (4-bit) quantization while maintaining near-baseline accuracy. The model handles text, image, and video inputs with a 262K context window. It is Apache 2.0 licensed, gated-free, and optimized for vLLM inference on NVIDIA GPUs. Primary use cases include AI agents, chatbots, RAG systems, and conversational applications.

Source: HuggingFace — huggingface.co/nvidia/Qwen3.5-122B-A10B-NVFP4
64.6B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
89.4k
Downloads (30d)

Key facts

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

FieldValue
Developernvidia
Parameters64.6B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads89.4k
Likes33
Last updated2026-06-02
Sourcenvidia/Qwen3.5-122B-A10B-NVFP4

What Qwen3.5-122B-A10B-NVFP4 is

A post-training quantized derivative of Qwen3.5-122B-A10B using NVIDIA Model Optimizer v0.44.0. Quantizes weights and activations of linear operators in transformer blocks to NVFP4 (4-bit floating point), reducing parameters from 16-bit to 4-bit representation. Activates 10B of 122B parameters during inference (MoE architecture). Calibrated on CNN DailyMail and Nemotron-Post-Training-Dataset-v2. Evaluated on MMMU Pro, GPQA Diamond, SciCode, AA-LCR, and IFBench with minimal accuracy loss versus FP8 baseline (e.g., GPQA Diamond 87.37% → 86.77%). Requires vLLM for inference, tested on NVIDIA B200 GPU. Context length stated as up to 262K; official model card references this property.

Quickstart

Run Qwen3.5-122B-A10B-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/Qwen3.5-122B-A10B-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

AI Agent and Chatbot Deployment

The model's instruction-following (IFBench 70.80%), tool-choice capabilities, and conversational design make it suitable for building autonomous agents and interactive chatbots. MoE architecture with 10B active parameters reduces latency versus dense 122B models. vLLM integration enables fast batch serving.

Retrieval-Augmented Generation (RAG) Systems

Strong long-context performance (AA-LCR 67.13%) and 262K context window support efficient document ranking and in-context retrieval. NVFP4 quantization reduces GPU VRAM footprint, lowering infrastructure costs for multi-document RAG pipelines.

Scientific and Technical Code Analysis

Evaluated on SciCode (41.79%) and GPQA Diamond (86.77%); suitable for code generation, scientific reasoning, and domain-specific problem-solving. Multimodal input (images, diagrams) enhances utility for technical documentation and research workflows.

Running & fine-tuning it

ESTIMATE: ~30–31 GB GPU VRAM (122B parameters at 4-bit ≈ 61 GB unquantized; 4x reduction ≈ 15–16 GB model weights + 10–15 GB KV cache and activations depending on context length and batch size). Tested on NVIDIA B200 (141 GB total VRAM). Suitable for A100 (80 GB) with careful batching; H100 (80 GB) recommended for production. Requires CUDA 12.x support and vLLM compatible PyTorch. Operating system: Linux preferred per card.

Not clearly stated in model card. NVFP4 quantization targets inference; fine-tuning feasibility unknown. QLoRA compatibility with 4-bit quantized weights is plausible but unsupported and untested. Recommend: (1) Fine-tune on unquantized Qwen3.5-122B-A10B base model if task-specific adaptation is critical, then quantize with Model Optimizer. (2) Test LoRA adapters with quantized weights only after internal validation. (3) Refer to NVIDIA Model Optimizer documentation for quantization-aware training guidance.

When to avoid it — and what to weigh

  • Extreme Low-Latency Inference (<50ms) — MoE routing and vLLM-specific overhead may not meet hard real-time SLAs. Test end-to-end latency on target hardware; quantization does not guarantee speed without proper kernel optimization.
  • Specialized Domain Fine-Tuning without Validation — Base model exhibits known biases and toxic language from training data. Requires domain-specific evaluation and mitigation before deployment in sensitive industries (healthcare, legal, finance). NVIDIA recommends use-case-specific testing per V-model methodology.
  • Systems Without NVIDIA GPU Infrastructure — Model is optimized exclusively for NVIDIA GPU-accelerated systems (Blackwell confirmed; vLLM primary runtime). CPU-only or non-NVIDIA GPU deployment is not documented or supported.
  • Applications Requiring Guaranteed Factual Accuracy — Model card explicitly warns of potential inaccuracy, key information omission, and irrelevant/redundant text generation. Not suitable for high-stakes factuality without external validation (e.g., RAG with trusted corpora).

License & commercial use

Apache License 2.0. Standard permissive open-source license allowing modification, distribution, and private use without restriction. Model card explicitly states: 'This model is ready for commercial/non-commercial use.' No proprietary restrictions on the quantized artifact itself.

Apache 2.0 is an OSI-approved permissive license. Commercial use is explicitly permitted in the model card ('ready for commercial/non-commercial use'). No gating or commercial license required. However, responsibility for compliance with downstream data privacy, IP, and regulatory requirements (HIPAA, GDPR, etc.) remains with the deployer. NVIDIA disclaims liability and recommends internal review for industry-specific legal and ethical standards. Derivative works must retain Apache 2.0 attribution.

DEV.co evaluation signals

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

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

Model card does not claim or assess security posture. Key considerations: (1) Quantized weights and activations may introduce side-channel attack surface if deployed in shared GPU environments; vLLM isolation is model-responsibility, not kernel-guaranteed. (2) Model trained on internet-crawled data; may encode backdoors, adversarial patterns, or prompt injection vulnerabilities. (3) Multimodal inputs (image, video) expand attack surface; model card warns to 'ensure proper rights and permissions' but does not detail content filtering. (4) Activates only 10B of 122B parameters (MoE); routing logic security is not documented. (5) No security audit, formal verification, or penetration testing results provided. NVIDIA maintains a vulnerability reporting channel (app.intigriti.com/programs/nvidia); use it for discovered issues. Standard LLM safety practices apply: input validation, output filtering, rate limiting.

Alternatives to consider

Qwen/Qwen3.5-122B-A10B-FP8

Unquantized Alibaba baseline with higher accuracy (GPQA 87.37% vs. 86.77%) but ~2x larger model size (~122B FP8 ≈ 244 GB). Choose if accuracy is critical and VRAM is abundant; otherwise NVFP4 offers better efficiency trade-off.

Llama 3.1 405B (Meta)

Denser 405B model; more parameters than Qwen3.5-122B-A10B but larger VRAM footprint. Strong on benchmarks; different training corpus and evaluation methodology. License (LLAMA 2 Community): requires review for commercial use beyond Meta's allowances. Choose if native English instruction-following is priority; Qwen excels on code and tool use.

Mixtral 8x22B (Mistral)

Smaller open-weight MoE model (~141B total, ~39B active); lower VRAM requirement. Apache 2.0 license (clear commercial use). Weaker on benchmarks (GPQA, code) than Qwen3.5; suitable for cost-optimized RAG and general chat. No official NVFP4 quantization; use GPTQ or AWQ instead.

Software development agency

Ship Qwen3.5-122B-A10B-NVFP4 with senior software developers

Qwen3.5-122B-A10B-NVFP4 offers efficient inference on NVIDIA GPUs with minimal accuracy trade-off. Test it in your environment with vLLM, validate on your benchmark tasks, and contact our team to architect a robust RAG or AI agent deployment. Learn more about quantization strategies and production best practices.

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.

Qwen3.5-122B-A10B-NVFP4 FAQ

Can I use this model commercially without additional licensing or NVIDIA approval?
Yes. The model is Apache 2.0 licensed and the model card explicitly states it is 'ready for commercial/non-commercial use.' No commercial license, approval, or royalty is required. However, you are responsible for ensuring compliance with data privacy laws (GDPR, CCPA, etc.), IP rights for training data and inference inputs, and regulatory requirements in your industry (HIPAA for healthcare, etc.). NVIDIA disclaims liability for downstream misuse.
What GPU hardware do I need to run this model?
Estimated ~30–31 GB GPU VRAM for inference. Suitable for NVIDIA A100 (80 GB), H100 (80 GB), H200 (141 GB), or B200 (141 GB). The model was tested on B200. vLLM is the primary serving framework; requires CUDA 12.x and Linux OS. CPU-only inference is not supported. For production, H100 or B200 is recommended to support batch inference and long contexts.
How much accuracy loss is there due to NVFP4 quantization compared to the base FP8 model?
Minimal. On benchmark averages: GPQA Diamond 87.37% (FP8) → 86.77% (NVFP4) = 0.60% drop. SciCode 42.16% → 41.79% = 0.37% drop. IFBench 70.91% → 70.80% = 0.11% drop. AA-LCR actually improves: 65.5% → 67.13%. Trade-off is ~4x model size reduction (15–16 GB vs. 60–61 GB) for <1% accuracy loss on these benchmarks. Domain-specific accuracy on your data may differ; test before production.
Does this model support fine-tuning or LoRA adapters?
Not explicitly documented. NVFP4 quantization is inference-optimized. Official fine-tuning guidance is not provided. Recommendation: (1) If you need task-specific adaptation, fine-tune on the unquantized Qwen3.5-122B-A10B base model, then quantize with NVIDIA Model Optimizer. (2) LoRA adapter compatibility with quantized weights is untested; proceed with caution and validate on your workload. (3) Refer to NVIDIA Model Optimizer docs or open an issue on GitHub for official support.

Software development & web development with DEV.co

Adopting Qwen3.5-122B-A10B-NVFP4 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 LLM at scale?

Qwen3.5-122B-A10B-NVFP4 offers efficient inference on NVIDIA GPUs with minimal accuracy trade-off. Test it in your environment with vLLM, validate on your benchmark tasks, and contact our team to architect a robust RAG or AI agent deployment. Learn more about quantization strategies and production best practices.