Qwen3 72B Embiggened needs about 50.6 GB of memory at Q4_K_M (a 47.4 GB file) with an 8K context window, which is more than a 48 GB GPU holds. This repository has 25 quants, from 23.7 GB (IQ1_M) to 77.3 GB (Q8_0). Longer context needs more: at 32K tokens the KV cache alone adds about 10.7 GB.
Qwen3 72B Embiggened – a 72.7B‑parameter experimental model
Qwen3 72B Embiggened expands the Qwen3‑32B architecture to a full 72.7B‑parameter model using interpolation and layer duplication. It retains the Qwen3 group‑query attention design and is available in multiple GGUF quantizations.
Good at
- Large‑context reasoning up to 40960 tokens
- High‑capacity text generation
- Experimental scaling for future distillation
Limits
- Experimental model, performance may differ from native 72B
- No vision or multimodal support
- Requires large VRAM for full precision
Best for
- Research on model scaling
- Testing interpolation techniques
- Preparing for Qwen3‑235B distillation
Who it’s for. Researchers and developers interested in large‑scale language models and experimental interpolation methods. Suitable for those who need a 72B‑parameter architecture without training from scratch.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| IQ1_M | 23.7 GB | 26.9 GB | Too big | Get file → | |
| IQ2_XXS | 25.5 GB | 28.7 GB | Too big | Get file → | |
| IQ2_XS | 27.1 GB | 30.2 GB | Too big | Get file → | |
| IQ2_S | 27.9 GB | 31.1 GB | Too big | Get file → | |
| IQ2_M | 29.3 GB | 32.5 GB | Too big | Get file → | |
| Q2_K | 29.7 GB | 32.9 GB | Too big | Get file → | |
| Q2_K_L | 30.9 GB | 34.1 GB | Too big | Get file → | |
| IQ3_XXS | 31.8 GB | 34.9 GB | Too big | Get file → | |
| IQ3_XS | 32.8 GB | 35.9 GB | Too big | Get file → | |
| Q3_K_S | 34.3 GB | 37.5 GB | Too big | Get file → | |
| IQ3_M | 35.4 GB | 38.6 GB | Too big | Get file → | |
| Q3_K_M | 37.6 GB | 40.8 GB | Too big | Get file → | |
| Q3_K_L | 39.5 GB | 42.7 GB | Too big | Get file → | |
| IQ4_XS | 39.7 GB | 42.9 GB | Too big | Get file → | |
| Q3_K_XL | 40.6 GB | 43.8 GB | Too big | Get file → | |
| IQ4_NL | 41.3 GB | 44.5 GB | Too big | Get file → | |
| Q4_0 | 41.4 GB | 44.6 GB | Too big | Get file → | |
| Q4_K_S | 43.8 GB | 47.0 GB | Too big | Get file → | |
| Q4_1 | 45.7 GB | 48.9 GB | Too big | Get file → | |
| Q4_K_M | 47.4 GB | 50.6 GB | Too big | Get file → | |
| Q4_K_L | 48.3 GB | 51.5 GB | Too big | Get file → | |
| Q5_K_S | 51.4 GB (2 parts) | 54.6 GB | Too big | Get file → | |
| Q5_K_M | 54.4 GB (2 parts) | 57.6 GB | Too big | Get file → | |
| Q6_K | 64.3 GB (2 parts) | 67.5 GB | Too big | Get file → | |
| Q8_0 | 77.3 GB (2 parts) | 80.4 GB | Too big | Get file → |
For 8 GB VRAM, use IQ1_M (23.74 GB file, ~26.9 GB memory at 8 k context). With 16 GB VRAM, IQ2_S (27.93 GB file, ~31.1 GB memory) or Q2_K (29.72 GB file, ~32.9 GB memory). For 32 GB VRAM, Q4_K_M (47.37 GB file, ~50.6 GB memory) offers a good balance of size and quality.
What GPU do you need for Qwen3 72B Embiggened?
These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.
- 4 GB VRAM: no quant fits at an 8K context.
- 6 GB VRAM: no quant fits at an 8K context.
- 8 GB VRAM: no quant fits at an 8K context.
- 12 GB VRAM: no quant fits at an 8K context.
- 16 GB VRAM: no quant fits at an 8K context.
- 24 GB VRAM: no quant fits at an 8K context.
- 32 GB VRAM: the best quant that fits is IQ2_S (27.9 GB file, about 31.1 GB in use).
- 48 GB VRAM: the best quant that fits is Q4_K_S (43.8 GB file, about 47.0 GB in use).
Which Qwen3 72B Embiggened quant should you choose?
Q4_K_M is the usual starting point for Qwen3 72B Embiggened: it balances file size and quality well. Here it is a 47.4 GB file.
The Q8_0 file is 77.3 GB, about 1.6 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.
If memory is tight, the smallest file is IQ1_M at 23.7 GB, with the quality trade-offs that come with fewer bits. Use it only when memory is very tight. Expect a clear drop in quality, which is usually worse on small models and easier to live with on large ones.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about Qwen3 72B Embiggened
How much VRAM does Qwen3 72B Embiggened need?
About 50.6 GB at Q4_K_M with an 8K context window. That is the 47.4 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.
Can I run Qwen3 72B Embiggened on an 8 GB GPU?
Not at an 8K context. The smallest file needs about 26.9 GB, so you would need a larger GPU or CPU offloading.
What is the file size of Qwen3 72B Embiggened GGUF?
It depends on the quant: IQ1_M is 23.7 GB, IQ2_XXS is 25.5 GB, IQ2_XS is 27.1 GB, IQ2_S is 27.9 GB, IQ2_M is 29.3 GB, Q2_K is 29.7 GB and 19 more.
What context length does Qwen3 72B Embiggened support?
The GGUF header lists a trained context length of 40,960 tokens (40K). The KV cache grows with the context you use: about 2.7 GB at 8K and 10.7 GB at 32K.
Where can I download Qwen3 72B Embiggened GGUF?
The files are on Hugging Face in the bartowski/cognitivecomputations_Qwen3-72B-Embiggened-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.
What license does Qwen3 72B Embiggened use?
Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.
Can I run Qwen3 72B Embiggened on 8 GB VRAM?
Use the IQ1_M quant, which requires about 26.9 GB of VRAM at 8 k context. It is the smallest quant available.
Does this model support vision or multimodal inputs?
No, the model only handles text. It has no vision or multimodal capabilities.
Is the model fully trained?
It is experimental; created via interpolation and duplication, not further trained. Performance may differ from a native 72B model.