My Qwen 3.8 27B tests on limited VRAM (16-20GB)

I have done several tests of different versions of Qwen 3.8 27B, mainly IQ4_XS since this is what fits my card. Some other models for comparison. Test 1 Create a single-file Three.js animation: a teal-and-gold mechanical flower unfolds around a glowing core, then closes seamlessly. Cinematic lighting, visible gears, orbiting camera, no text. This is a hard task taken from a test from frontier models and even these weren't perfect. No model scored 5/5, which is to be expected. Test 2 This is a simple Tauri+Yew desktop app that is supposed to display a rich text editor. The editor should be a component so it is easy to reuse it in another app. AGENTS.md file: # Functionality - HTML mode - WYSIWYG mode ## BB Code tags - `<b></b>`: bold text - `<i></i>`: italic text - `<u></u>`: underline - `<span style="color: #FFAA41;"></span>`: text color - `<sub></sub>`: subscript - `<sup></sup>`: superscript - `<h1></h1>`: header (h1 to h6) - `<pre><code></code></pre>`: code block - `<blockquote></blockquote>`: quotation block - `<a href="https://"></a>`: link with a prompt for the link and link text - handling `&nbsp;`, `&lt;`, `&gt;` - mouseover tooltips There should be a possibility of combining two or more, e.g. bold red text. Editing should be possible for both modes. Test 3 Give me 20 Polish words with 3 letters Funny, but some do get stuck on this one. I suspect quantization-related issues. I'd keep this test to test quantizations then. Results: For test 1 GSQ-RCO IQ3_S and Byteshape versions did the best job, although each had an issue with proper lightning. Twin Turbo screwed up gears, but petals are better than Unsloth, which can get extra scores for shadows. Cold Fusion totally botched it, Nex N2 Mini (APEX Quality) botched it completely. Swift model did provide something close to Unsloth. Test 2 which is a long complex task - Unsloth seems most reliable delivering a working solution in an hour. - Twin Turbo did it in 15 minutes, but with minor bugs that took it 10 extra minutes to solve, so it can be recommended as well. May require a bit more detailed agents file describing the tests or so. I also asked it to add line numbers to the WYSIWYG mode which took another 25 minutes and extra prompts since it is kind of complex. Needs to account for variable font sizes. - Swift model took several hours, stopped several times for no reason, but delivered, although with minor bugs that it fixed - GSQ-RCO delivered but took 5h to do so, stopped once, but as with swift I just needed to tell it 'continue' - Nex N2 Mini: major problems with basic functionality, do not recommend - Byteshape: minor issues after 4h work, but it fixed them - HauhauCS Aggressive: color picker issues after several hours, but it can deliver Test 3 which was taken from Reddit, surprisingly Unsloth does a poor job here: - Unsloth: stuck in a loop of repeating 1 word - Nex N2 Mini: stuck - Twin Turbo: goes for a while, but delivers - Byteshape: gives a proper result fast - Swift: like Byteshape - GSQ-RCO: works fine - Gemma4 12B Q5_K_M: instant result Swift model takes a significantly larger memory footprint than other IQ4_XS version which allows only for 128k context window on 7900XT, so I'd only consider it in Q4_K_M version on a 24GB card. Other versions allowed for at least around 150k easily. On a 16GB card GSQ-RCO seems to be the winner here with Byteshape being not far behind. Both take more time to do it, which can be considered a cost.