Qwen3.6 35B A3B heretic MTP IQ4 XS needs about 20.4 GB of memory at IQ4_XS (a 19.2 GB file) with an 8K context window, so it fits on a 24 GB GPU. Longer context needs more: at 32K tokens the KV cache alone adds about 2.8 GB.
Qwen3.6 35B A3B heretic MTP IQ4_XS – a decensored, MTP‑preserved model for single‑GPU use
This Qwen3.6 35B model is quantized to IQ4_XS and retains 19 MTP tensors for speculative decoding. It is designed for single‑GPU setups with 12 GB VRAM, offloading part of the experts to system RAM. The model reduces refusals while keeping close to the original quality.
Good at
- Reasoning
- Tool usage
- MTP‑preserved speculative decoding
Limits
- Requires 12 GB VRAM with 32 GB system RAM
- Large memory footprint – 20.4 GB at 8 k context
- No vision support
Best for
- Single‑GPU inference on 12 GB cards
- Speculative decoding with experts offloaded
- Reducing refusal rates in applications
Who it’s for. Users needing a decensored Qwen3.6 model that can run on a single 12 GB GPU while still supporting MTP for speculative decoding. It is suitable for developers who want lower refusal rates without requiring multi‑GPU setups.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| IQ4_XS | 19.2 GB | 20.4 GB | Too big | Get file → |
The IQ4_XS quantization results in a 19.18 GB file and requires about 20.4 GB of VRAM at 8 k context. For a 12 GB GPU, keep 24 of 40 expert layers in system RAM. If you have 8 GB VRAM, the model will not fit without further quantization or offloading.
What GPU do you need for Qwen3.6 35B A3B heretic MTP IQ4 XS?
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: the best quant that fits is IQ4_XS (19.2 GB file, about 20.4 GB in use).
Which Qwen3.6 35B A3B heretic MTP IQ4 XS quant should you choose?
IQ4_XS (19.2 GB) is the middle option in this repository. The usual sweet spot between file size and quality, and the quant most people start with.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Frequently asked questions about Qwen3.6 35B A3B heretic MTP IQ4 XS
How much VRAM does Qwen3.6 35B A3B heretic MTP IQ4 XS need?
About 20.4 GB at IQ4_XS with an 8K context window. That is the 19.2 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 Qwen3.6 35B A3B heretic MTP IQ4 XS on an 8 GB GPU?
Not at an 8K context. The smallest file needs about 20.4 GB, so you would need a larger GPU or CPU offloading.
What is the file size of Qwen3.6 35B A3B heretic MTP IQ4 XS GGUF?
It depends on the quant: IQ4_XS is 19.2 GB.
What context length does Qwen3.6 35B A3B heretic MTP IQ4 XS support?
The GGUF header lists a trained context length of 262,144 tokens (256K). The KV cache grows with the context you use: about 0.7 GB at 8K and 2.8 GB at 32K.
Where can I download Qwen3.6 35B A3B heretic MTP IQ4 XS GGUF?
The files are on Hugging Face in the hermeticvm/Qwen3.6-35B-A3B-heretic-MTP-IQ4_XS-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 Qwen3.6 35B A3B heretic MTP IQ4 XS 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 this model on 8 GB VRAM?
No, the model requires 12 GB VRAM with part of the experts in system RAM. 8 GB is insufficient.
Does this model support vision?
No, it does not include vision capabilities.
What license does it use?
The model is released under the apache-2.0 license.