Qwen-4B Bugio: Qwen3.8-distilled 4B model with faster thinking and vision for 4 GB GPUs (GGUF)

Qwen-4B Bugio: a tiny local AI tuned on Qwen3.8 reasoning traces, with thinking and vision

Qwen-4B Bugio (Bugio-4B) is a tiny, experimental open-weights model for Ollama and llama.cpp. It is a 2.4 GB build of Qwen3.8-4B-Distill, a Qwen3.5-4B fine-tuned on reasoning traces from Qwen3.8. It thinks step by step with shorter reasoning, reads images and screenshots, and calls tools, all offline. It runs on a 4 GB GPU from 2014.

ollama run hf.co/sysrobson/Qwen-4B-Bugio-Vision-Thinking-GGUF:IQ3_XXS

Experimental. Bugio is the reasoning sibling of Sagui-4B, built with the same recipe and the same size. On our reasoning questions it used about 1.9ร— fewer thinking tokens than Sagui, but it is weaker at TypeScript. For coding, use Sagui.

Support Bugio and Sagui: donate via PayPal

Bugio and Sagui are free. Donations pay for the compute behind the next tiny models: โ˜• PayPal

Bugio or Sagui?

Both are 2.4 GB, use the same compression recipe, and were measured on the same 4 GB GTX 970.

Bugio-4B (this model) Sagui-4B
Source weights Qwen3.8-4B-Distill (Qwen3.5-4B tuned on Qwen3.8 traces) Qwen3.5-4B
Average thinking length, 5 reasoning questions 1447 tokens, 101 s 2786 tokens, 198 s
Correct answers on those questions 5/5 5/5
HumanEval-TS pass@1 36.5% 44.0%
MBPP-TS pass@1 46.2% 54.6%
Vision, thinking, tool calling โœ… โœ…
Total size with vision 2.38 GB 2.38 GB
Best for Quick reasoning, Q&A, chat, screenshots TypeScript and code

A bugio is a howler monkey, native to Brazil's Atlantic Forest. Its call can be heard kilometres away. Small, loud and quick to answer.

See it working

These are unedited outputs from this exact GGUF. Prompts and raw responses are in examples/transcripts.json, and scripts/run_demos.py reproduces them. Sampling isn't seeded, so your wording will vary.

Thinking mode

Prompt: A train leaves at 14:35 and arrives at 17:10. How long is the trip in minutes?

The trip takes **155 minutes**.

**Breakdown:**
- From 14:35 to 17:10 is **2 hours and 35 minutes**.
- 2 hours = 120 minutes
- 120 + 35 = **155 minutes**

338 tokens, 24.3 s on a GTX 970, reasoning included. Sagui-4B spent 3,003 tokens and 212 s on the same question in our comparison run.

UI screenshot โ†’ working React component

Input screenshot Bugio's code, rendered
Original sign-up form Sign-up form rendered from the code Bugio wrote

Thinking on, 1545 tokens, 106.9 s on a GTX 970. Only the import line was removed to render it in the browser. It got the text, fields, button and link right, but missed the placeholders and centered the labels.

Screenshot โ†’ bug fix

Editor screenshot with a TypeScript TS2322 error

Prompt: This is a screenshot from my editor. Explain the error in one sentence, then give the corrected line 4.

The error occurs because the reduce callback's second parameter i is the accumulator object from the previous iteration, which is a number, but the code tries to access i.qty without a type assertion, causing a type mismatch.

return items.reduce((sum, i) => sum + i.price * i.qty, 0)

The fix is to add the type assertion : number to the second parameter i of the reduce callback, so TypeScript correctly infers that i is an Item object.

120 tokens, 30.9 s on a GTX 970. It read the code from the image and the corrected line is right. The explanation is wrong: the real cause is the string initial value "0". Sagui explained this one correctly.

Tool calling

Asked "What's the weather in Lisbon? Use the tool.", Bugio called get_weather with {"city": "Lisbon"}. Given the tool result, it answered: "The weather in Lisbon is currently 21ยฐC and sunny."

Run it

Ollama, one command

ollama run hf.co/sysrobson/Qwen-4B-Bugio-Vision-Thinking-GGUF:IQ3_XXS

Ollama, from the files

hf download sysrobson/Qwen-4B-Bugio-Vision-Thinking-GGUF --local-dir bugio
cd bugio
ollama create bugio-4b -f Modelfile
ollama run bugio-4b "What is in this image? ./photo.png"

Control thinking per chat with --think or --think=false, or in the API:

curl http://localhost:11434/api/chat -d '{
  "model": "bugio-4b",
  "think": true,
  "messages": [{"role": "user", "content": "A shirt costs $40 after a 20% discount. What was the original price?"}]
}'

llama.cpp

llama-server -m Bugio-4B-IQ3_XXS.gguf --mmproj mmproj-Bugio-4B-F16.gguf -ngl 99 -c 16384 --temp 0.6 --top-p 0.95 --top-k 20

Don't use greedy decoding (temperature 0): the source model can loop. The Modelfile already sets the sampling recommended by its authors.

Tip for 4 GB GPUs

Ollama can split the model between CPU and GPU to be safe. On our GTX 970 that gave 15.5 tok/s. Forcing every layer onto the GPU gave 23.2 tok/s:

PARAMETER num_gpu 99

Add that line to the Modelfile and run ollama create again. If you run out of memory, remove it or lower num_ctx.

Files

File Size Needed for
Bugio-4B-IQ3_XXS.gguf 1.71 GB Everything
mmproj-Bugio-4B-F16.gguf 0.67 GB Images only
Modelfile, template, params small Ollama setup

Benchmarks

Reasoning: thinking length

Five short arithmetic and logic questions with thinking on, through Ollama with the Modelfile settings. Questions, answers and timings are in benchmarks/thinking.json, and scripts/think_compare.py reproduces them. Five questions show the size of the gap, not a precise accuracy figure.

Bugio-4B Sagui-4B
Correct 5/5 5/5
Average tokens per answer 1447 2786
Average time per answer, GTX 970 101 s 198 s

Code: TypeScript

TypeScript code generation on MultiPL-E with greedy decoding. A solution passes only if it type-checks with tsc and passes its tests in Node.

Benchmark Bugio-4B IQ3_XXS Sagui-4B IQ3_XXS
HumanEval-TS 58/159 (36.5%) 70/159 (44.0%)
MBPP-TS 180/390 (46.2%) 213/390 (54.6%)
Weighted average 43.4% 51.5%

These measure raw completion without thinking mode. Most of Bugio's extra failures are tsc type errors. Completions, failures and the script are in benchmarks/.

Measured on an Intel i5-10400, 16 GB RAM and an NVIDIA GTX 970 (4 GB). llama.cpp benchmark speed: 21.3 tok/s. CPU only through Ollama: 9.1 tok/s.

Limitations

  • Experimental and weaker at code. It scores about 8.1 points below Sagui-4B on TypeScript. In our groupBy demo it reused the parameter name key for a local variable, which doesn't compile. Use Sagui for code.
  • Unverified upstream data. Qwen3.8-4B-Distill was trained on about 45,000 reasoning traces that its authors describe but haven't published. The authors also report a lower GSM8K score than the base Qwen3.5-4B and a higher MMLU score.
  • Latin scripts only. Removing other alphabets is part of how the model got this small. Use the original Qwen3.5-4B for Chinese, Japanese, Korean, Russian, Arabic and similar languages.
  • Heavy quantization. IQ3_XXS trades accuracy for size, and it was calibrated on TypeScript.
  • Vision is inherited, not tuned. The projector is Qwen3.5-4B's, unmodified. It read screenshots well in our tests but explained one of them wrongly. It was not tested on photos at scale.
  • It makes mistakes. Review its answers and code, and don't use it for medical, legal or safety-critical decisions.

FAQ

Is there a Qwen3.8 4B model? Not an official one. Qwen released Qwen3.8 as a 27B dense model and a large MoE. Bugio is built on Qwen3.8-4B-Distill, a community fine-tune of Qwen3.5-4B on Qwen3.8 reasoning traces. It is not Qwen3.8 itself.

Can I run Qwen3.8 reasoning on a 4 GB GPU? This is the closest option we know of. Bugio fits in 2.4 GB with vision, and every result here comes from a 4 GB GTX 970.

Is this Qwen 4? No. "Qwen-4B" means a Qwen-family model with 4 billion parameters.

Bugio or Sagui? Bugio answered our reasoning questions about twice as fast, with shorter thinking. Some questions still trigger long reasoning. Sagui is clearly better at TypeScript. Both see images and call tools.

Does it run without a GPU? Yes, on CPU, at 9.1 tok/s on an i5-10400.

Was it retrained? Not by us. It has the same weights as Qwen3.8-4B-Distill, with a trimmed vocabulary and heavier quantization. See the recipe.

Can I use it commercially? Yes, under Apache 2.0. Keep the LICENSE and NOTICE files.

Em portuguรชs

Qwen-4B Bugio รฉ uma IA local experimental de 2,4 GB, ajustada com raciocรญnios do Qwen3.8, que pensa e enxerga. Roda offline com Ollama ou llama.cpp, inclusive em GPU de 4 GB. ร‰ o irmรฃo do Sagui-4B, com a mesma receita e o mesmo tamanho. Nas nossas perguntas de raciocรญnio, pensou cerca de metade do que o Sagui pensa antes de responder, mas รฉ mais fraco em TypeScript. Para programar, use o Sagui.

ollama run hf.co/sysrobson/Qwen-4B-Bugio-Vision-Thinking-GGUF:IQ3_XXS

Nรฃo รฉ um modelo oficial do Qwen3.8: รฉ uma versรฃo compactada do Qwen3.8-4B-Distill, um Qwen3.5-4B ajustado com raciocรญnios gerados pelo Qwen3.8.

Gostou? Apoie o projeto. As doaรงรตes pagam a computaรงรฃo dos prรณximos modelos pequenos e gratuitos.

Support Bugio and Sagui: donate via PayPal

Support the project

Bugio and Sagui are free and stay free. Donations pay for the compute and the time behind the next tiny models.

Support Bugio and Sagui: donate via PayPal

No money to spare? A โค๏ธ on this page and a post about what you built help just as much.

Credits and license

Bugio-4B is a modified version of Qwen3.8-4B-Distill by empero-ai, itself a fine-tune of Qwen3.5-4B by the Qwen team. It was quantized with llama.cpp and is licensed under Apache 2.0. The changes are listed in NOTICE. This is an independent community project, not affiliated with or endorsed by the Qwen team, Alibaba Cloud or empero-ai.

Downloads last month
282
GGUF
Model size
4B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

3-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for sysrobson/Qwen-4B-Bugio-Vision-Thinking-GGUF

Finetuned
Qwen/Qwen3.5-4B
Quantized
(32)
this model

Evaluation results