Qwopus3.8 -27B Flash Review, Image Recognition From a HouseLLM!

Jack Rongs re-tuned models are showing again and again absolutely incredible progression in the LLM retuned space. In this case Qwen 3.8 debuted with very high benchmarks, then Jack took them farther! Not only is this a SOTA (State of the Art) houseLLM that can run on minimal equipment - it's free, and can do image recognition! Powerful! For you interest here are some basic performance benchmarks, but please note - this is a slow and very intelligent 35B sized model:

• TerminalBench 2.1: 73.0
• SWE-bench Pro: 61.7
• QwenSWEBench: 79.0
• IFBench: 79.5
• GPQA Diamond: 89.2
• HLE: 30.8

What is a Re-tuned Model?

It is simply this:

A retuned LLM is a pretrained language model that has been further adjusted so its behavior changes for a particular purpose—such as better coding, scientific reasoning, instruction-following, or a specific writing style. This is typically done through additional fine-tuning, preference training, or modifying how the model is prompted and configured, while retaining most of the knowledge learned during its original training.

Downloading and Setting up a Image Capable LLM.

1. Get the main model: On the right side of the page you will see your model selection. Typically - and this is subject to ongoing change - to get a reliable model you will want something that is atypically a minimum of 4-bit, however progress is accelerating quickly into reliable and useful 2-bit quants. Quants are basically recompression of the model weights to fit on smaller equipment.  

2. You will need the mmproj-F32.gguf. The right side menu may throw you off it's at the top under the file - as in:

Clicking on that file will take you to another menu - where you will finally get to download it:

3. Get a Running llm-server. So you can try a quick download and build as in:

git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp && cmake -B build && cmake --build build --config Release -t llama-server -j

When it is done it just puts it in a  local directory so you will need to copy it's contents to /usr/bin

./build/bin/llama-server

If you want a very detailed guide of making a very powerful llm-server build with TurboQuant KV-Caching consider a this good walk-through guide. Even though it's a 'StudentLLM' it is a full power-build that will run any model.

StudentLLM - Qwen2.5-coder-7b-instruct-q6-k / Qwen3.5 Agentic on a Ryzen 5-2600/ 3060ti. Production LLM or not? YES!

We Look a StudentLLM setup to get as much productivity out of limited hardware as we can.

Hot Configthinkmeltprotonmail.com

In our case our minimal run script required a fair bit of tweaking, and this is how we were able to get it to work.

• We were forced to make sure the mmproj file was not offloaded to GPU as this was breaking stuff, so our valid run configuration became:

/usr/bin/llama-server --jinja \
-m /home/c/models/Qwopus3.8-27B-Flash.i1-Q6_K.gguf \
--mmproj /home/c/models/mmproj-F32.gguf \
-ngl auto \
--no-mmproj-offload \
--host 192.168.1.3 \
--port 8080

One will notice we absolute pathed everything - it just works better out the gate. Instead of llama-server use /usr/bin/llama-server.  

Managing Expectations

This stuff is awesome, but no it's not going to be instant or replace your professional subscription for Claude, or ChatGPT - nor should it with their budgets that rival small nations, but the key is it's your LLM, and YOU now know how to use it. People in attorney-client privilege scenarios, doctors offices, privacy advocates simply do not want their data harvested, and not everybody wants to commit the $40 - $400/month to serious grade-llms. Additionally this runs on about $2000 in hardware.  That is where this site specializes in - small budgets, maximum performance on mini-bucks. You the small-guy are the winner!

Once it finally loads - your console back-end should look something like this:

First Run

We are not going to go too deeply into a standard LLM review as we have covered that a dozen times before, but will focus on the image-to-text aspects.

• This LLM is slow on our 4080 it came in at 4.8 T/s - ouch. Atypically it's been our experience that anything below 15 T/s is more of a test LLM - and something people do not have the patience to wait on.

Image Recognition was Powerful and Fast!

• This LLM generates  slow - but we literally screen shot a log fired the image back at it, and it nailed it immediately, wow!

It took it only a few seconds to cleanly identify a cat

Can It Watch For Burglars - Absolutely!

It is pretty smart, really.

Conclusion

This LLM could have many, many functions. The sky is the limit. Monitor your 3D printer to make sure that your print is good, monitoring a process to make sure that it is operating safely. Would it be safety certified we think not - but it definitely can be a ongoing security scanner, and the cost is free - that's powerful and you are the benefactor.