Qwen3.5 122B A10B NVFP4 — Hardware Requirements & GPU Compatibility
Chat
An NVFP4-quantized version of Alibaba's Qwen3.5 122B A10B, repackaged by txn545. This large mixture-of-experts model has 122 billion total parameters with roughly 10 billion activated per token, giving it approximately 64.4 billion effective parameters in its quantized form. The NVFP4 (NVIDIA FP4) format is designed specifically for NVIDIA GPUs with FP4 support, offering aggressive compression while leveraging hardware-level acceleration. Qwen3.5 represents a significant generational upgrade in Alibaba's model lineup, and the 122B A10B variant delivers strong reasoning, coding, and multilingual performance. Despite the model's large total parameter count, the sparse activation pattern combined with FP4 quantization makes it feasible to run on high-end consumer GPUs, though multi-GPU setups will provide the best experience for longer context lengths.
197.8K downloads
16 likesSpecifications
- Publisher
- txn545
- Family
- Qwen
- Parameters
- 64.4B
- Architecture
- Qwen3_5MoeForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 248,320
- Release Date
- 2026-03-01
- License
- Apache 2.0
Get Started
HuggingFace
How Much VRAM Does Qwen3.5 122B A10B NVFP4 Need?
Select a quantization to see compatible GPUs below.
| Quantization | Bits | VRAM | + Context | File Size | Quality |
|---|---|---|---|---|---|
| BF16 | 16.00 | 129.1 GB | 138.7 GB | 128.71 GB | Brain floating point 16 — preferred for training |
Which GPUs Can Run Qwen3.5 122B A10B NVFP4?
BF16 · 129.1 GBQwen3.5 122B A10B NVFP4 (BF16) requires 129.1 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 168+ GB is recommended. Using the full 262K context window can add up to 9.6 GB, bringing total usage to 138.7 GB. No single GPU has enough memory — multi-GPU or cluster setups are needed.
Which Devices Can Run Qwen3.5 122B A10B NVFP4?
BF16 · 129.1 GB4 devices with unified memory can run Qwen3.5 122B A10B NVFP4, including NVIDIA DGX H100, NVIDIA DGX A100 640GB, Mac Pro M2 Ultra (192 GB).
Runs great
— Plenty of headroom
Decent
— Enough memory, may be tightRelated Models
Frequently Asked Questions
- How much VRAM does Qwen3.5 122B A10B NVFP4 need?
Qwen3.5 122B A10B NVFP4 requires 129.1 GB of VRAM at BF16. Full 262K context adds up to 9.6 GB (138.7 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 64.4B × 16 bits ÷ 8 = 128.7 GB
KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)
KV Cache + Overhead ≈ 10 GB (at full 262K context)
VRAM usage by quantization
BF16129.1 GBBF16 + full context138.7 GB- Can NVIDIA GeForce RTX 5090 run Qwen3.5 122B A10B NVFP4?
No — Qwen3.5 122B A10B NVFP4 requires at least 129.1 GB at BF16, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
- Can I run Qwen3.5 122B A10B NVFP4 on a Mac?
Qwen3.5 122B A10B NVFP4 requires at least 129.1 GB at BF16, which exceeds the unified memory of most consumer Macs. You would need a Mac Studio or Mac Pro with a high-memory configuration.
- Can I run Qwen3.5 122B A10B NVFP4 locally?
Yes — Qwen3.5 122B A10B NVFP4 can run locally on consumer hardware. At BF16 quantization it needs 129.1 GB of VRAM. Popular tools include Ollama, LM Studio, and llama.cpp.
- How fast is Qwen3.5 122B A10B NVFP4?
At BF16, Qwen3.5 122B A10B NVFP4 can reach ~23 tok/s on AMD Instinct MI300X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: AMD Instinct MI300X → 5300 ÷ 129.1 × 0.55 = ~23 tok/s
Estimated speed at BF16 (129.1 GB)
AMD Instinct MI300X~23 tok/sReal-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
- What's the download size of Qwen3.5 122B A10B NVFP4?
At BF16, the download is about 128.71 GB.