Qwen3.5 4B MTP Heretic needs about 4.4 GB of memory at Q4_K_M (a 2.8 GB file) with an 8K context window, so it fits on a 6 GB GPU. This repository has 16 quants, from 0.0 GB to 3.6 GB (Q6_K). Longer context needs more: at 32K tokens the KV cache alone adds about 4.4 GB.
Qwen3.5 4B MTP Heretic – a low‑refusal, multi‑token‑prediction vision‑language model
Qwen3.5 4B MTP Heretic is a 4.33 billion‑parameter model that has had its refusal behaviour ablated using heretic v1.4.0 while preserving the multi‑token‑prediction head. It is a vision‑language model with a 262144‑token context window and is released under Apache‑2.0.
Good at
- Low‑refusal outputs
- Multi‑token prediction
- Vision‑language tasks
Limits
- No safety guarantees
- No default vision support listed
- Large VRAM requirement for higher quants
Best for
- Custom safety‑filtered deployments
- Vision‑language inference on local GPU
- Research on refusal ablation
Who it’s for. Developers and researchers who need a Qwen3.5‑4B variant with reduced refusal behaviour and an intact MTP head. The model is suitable for local inference on GPUs with sufficient VRAM and for projects requiring a vision‑language backbone.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| — | 0.0 GB | 1.6 GB | Fits | Get file → | |
| Q2_K | 2.0 GB | 3.6 GB | Fits | Get file → | |
| Q3_K_S | 2.1 GB | 3.7 GB | Fits | Get file → | |
| IQ3_S | 2.2 GB | 3.8 GB | Fits | Get file → | |
| IQ3_M | 2.2 GB | 3.8 GB | Fits | Get file → | |
| Q3_K_M | 2.3 GB | 3.9 GB | Fits | Get file → | |
| Q3_K_L | 2.5 GB | 4.1 GB | Fits | Get file → | |
| IQ4_XS | 2.6 GB | 4.2 GB | Fits | Get file → | |
| Q4_0 | 2.6 GB | 4.2 GB | Fits | Get file → | |
| Q4_K_S | 2.6 GB | 4.2 GB | Fits | Get file → | |
| IQ4_NL | 2.7 GB | 4.3 GB | Fits | Get file → | |
| Q4_K_M | 2.8 GB | 4.4 GB | Fits | Get file → | |
| Q4_1 | 2.8 GB | 4.4 GB | Fits | Get file → | |
| Q5_K_S | 3.1 GB | 4.7 GB | Fits | Get file → | |
| Q5_K_M | 3.2 GB | 4.8 GB | Fits | Get file → | |
| Q6_K | 3.6 GB | 5.2 GB | Fits | Get file → |
For 8 GB VRAM, use Q4_0 (file size 2.62 GB, estimated 4.2 GB memory). For 12 GB VRAM, Q5_K_S (3.07 GB, 4.7 GB memory) offers a balance of size and quality. For 16 GB VRAM, Q6_K (3.56 GB, 5.2 GB memory) provides the highest fidelity.
What GPU do you need for Qwen3.5 4B MTP Heretic?
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: the best quant that fits is Q3_K_M (2.3 GB file, about 3.9 GB in use).
- 6 GB VRAM: the best quant that fits is Q6_K (3.6 GB file, about 5.2 GB in use).
Which Qwen3.5 4B MTP Heretic quant should you choose?
Q4_K_M is the usual starting point for Qwen3.5 4B MTP Heretic: it balances file size and quality well. Here it is a 2.8 GB file.
If memory is tight, the smallest file is the first one in the table at 0.0 GB, with the quality trade-offs that come with fewer bits.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about Qwen3.5 4B MTP Heretic
How much VRAM does Qwen3.5 4B MTP Heretic need?
About 4.4 GB at Q4_K_M with an 8K context window. That is the 2.8 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 Qwen3.5 4B MTP Heretic on an 8 GB GPU?
Yes, with Q6_K: the 3.6 GB file needs about 5.2 GB at 8K context. Larger quants in this repository do not fit in 8 GB.
What is the file size of Qwen3.5 4B MTP Heretic GGUF?
It depends on the quant: file is 0.0 GB, Q2_K is 2.0 GB, Q3_K_S is 2.1 GB, IQ3_S is 2.2 GB, IQ3_M is 2.2 GB, Q3_K_M is 2.3 GB and 10 more.
What context length does Qwen3.5 4B MTP Heretic support?
The GGUF header lists a trained context length of 262,144 tokens (256K). The KV cache grows with the context you use: about 1.1 GB at 8K and 4.4 GB at 32K.
Where can I download Qwen3.5 4B MTP Heretic GGUF?
The files are on Hugging Face in the mradermacher/Qwen3.5-4B-MTP-Heretic-i1-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 Qwen3.5 4B MTP Heretic 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 Qwen3.5 4B MTP Heretic on 8 GB VRAM?
Yes, using the Q4_0 quant which requires approximately 4.2 GB of VRAM at 8 k context.
Does this model support vision tasks?
It is a vision‑language model, but no explicit vision files are listed; use the provided mmproj files if available.
What safety guarantees does it have?
The model has reduced safety filtering and no default safety guarantees; a policy layer is required for untrusted input.