Qwen3.5 4B Amadeus
Jitheswar’s abliterated build of Qwen3.5-4B: refusal behavior removed by directional ablation, with no fine-tuning. A 2.8 GB GGUF for 8 GB laptops. A community-modified checkpoint, not a Qwen release.
What do you want to do?
Choose a task. Get the software, files, settings, and steps to run it.
Open a guide for its software, hardware requirements, and setup steps.
At a glance
- Parameters
- 4B (Hugging Face lists about 5B total)
- Architecture
- Dense vision-language · community-modified checkpoint
- Context length
- 262,144 native (Qwen base model) · Jitheswar’s example uses 16,384
- License
- Apache 2.0, the same as the original model
- Software
- Ollama, llama.cpp, LM Studio, Transformers, vLLM, SGLang
Will it run on your machine?
Save your machine to see a personalized rating and its reasoning.
Add my hardwareBest for
- Private chat with fewer refusals on an entry-level laptop with 8 GB of GPU or unified memory.
- Try a local Hermes agent on modest hardware, as an experiment.
- Ask about images by loading the vision file alongside the model.
- Use Qwen3.5’s thinking mode, or turn it off for direct answers.
Tradeoffs
- Jitheswar reports 1 refusal on 304 test prompts, against 231 for the original model. That is the publisher’s own result, and a finite test does not guarantee every request is answered.
- Jitheswar reports slightly lower scores than the original on a truthfulness test (TruthfulQA 25.95% against 28.27%) and an instruction-following test (IFEval 78.19% against 81.70%).
- Fewer refusals do not make answers more accurate. Jitheswar’s card warns that answers on dangerous or specialized topics may be wrong or outdated.
- The card does not report tool-calling tests.
- Not an official Qwen release.
Ways to run it
Ollama · Windows, macOS, Linux
The smallest file, used by the Hermes guide on this page. Swap the tag for Q6_K for higher quality if you have memory to spare.
ollama pull hf.co/Jitheswar/Qwen3.5-4B-Amadeus:Q4_K_Mllama.cpp · Windows, macOS, Linux
Jitheswar’s example settings with the smaller file. Add --mmproj mmproj-Qwen3.5-4B-Amadeus-F16.gguf for image input, or --reasoning off to turn thinking off.
llama-server -hf Jitheswar/Qwen3.5-4B-Amadeus:Q4_K_M -ngl 99 -c 16384LM Studio · Windows, macOS, Linux
Search for Jitheswar/Qwen3.5-4B-Amadeus and download the Q4_K_M file.
Official instructionsvLLM · Linux GPU server
Full-precision safetensors; not used by the guide here.
vllm serve "Jitheswar/Qwen3.5-4B-Amadeus"What to expect
The pick for an entry-level laptop with 8 GB of GPU or unified memory. Start with the 2.8 GB Q4_K_M file in Ollama.
Explore its uses
Sources and files
Primary sources and files behind this model page.
Keep exploring
Bonsai 2 · 27B
A compressed 27B-class reasoning model with publisher-provided GGUF packs and a dedicated llama.cpp fork for CUDA, Metal, and CPU.
DeepSeek V4.1 Flash
An open-weight model for reasoning, coding, agents, and image understanding. DeepSeek V4.1 Flash combines a 552B backbone with sparse activation and up to one million tokens of context. Choose the original checkpoint or a community uncensored variant below.
Qwen3.8
Explore 4 versions of Qwen3.8. Compare model sizes, packages, hardware requirements, and available setups.