About this Resource
<p><i><span style="font-size: 16px;">Testing Ornith 1.5 35B to see how capable its Q4 and Q8 versions really are for local AI, coding, game development, 3D modeling, and more.</span></i></p><h2><br></h2><h2>Ornith 1.5 35B Is Surprisingly Good: Q4 vs Q8 Tested</h2><p><span style="font-size: 16px;">Ornith 1.5 35B-A3B is a new mixture-of-experts AI model built on top of Qwen 3.5. It is designed to offer stronger capabilities while still being practical to run on consumer hardware. The model is particularly interesting because there is currently no Qwen 3.8 35B-A3B equivalent available, making Ornith 1.5 35B a potentially useful alternative for people looking for a relatively compact model with strong coding and agent capabilities. The test compared two versions of the model, Q4_K_M and Q8 quantizations. Quantization is essentially a way of reducing a model's memory requirements so it can run on less powerful hardware. Q4 uses less memory, while Q8 generally preserves more of the model's quality.</span></p><h2><br></h2><h2>Q4 vs Q8: What Is the Difference?</h2><p><span style="font-size: 16px;">The Q4 version is the more practical option for users with limited hardware. It requires less memory and can run faster in many situations. The Q8 version uses significantly more memory, but the testing showed that it can produce noticeably better results on demanding tasks. Both versions were tested on the same system, using a 6000 Pro GPU, allowing their results to be compared directly.</span></p><h2><br></h2><h2>Strong Results in Browser and Game Creation</h2><p><span style="font-size: 16px;">One of the biggest strengths of Ornith 1.5 35B was its ability to create functional applications and games from instructions. In a browser-based operating system test, both versions created working interfaces with features such as applications, settings, games, notes, email and even a voice assistant. However, the Q8 version was clearly more capable. It produced a more detailed and polished operating system and was better at troubleshooting problems during development. The Q4 version still performed surprisingly well, especially considering its smaller memory requirements.</span></p><p><span style="font-size: 16px;"><br></span></p><h2>Q8 Was Better at Complex Coding Tasks</h2><p><span style="font-size: 16px;">The difference became more obvious in the Subway FPS test. The Q4 version created a basic 3D environment, but important controls such as mouse movement and shooting did not work correctly. The Q8 version initially had problems as well, but it was able to troubleshoot and fix many of them. After another round of improvements, movement, shooting and other gameplay features became functional. This highlights an important difference between the two versions: Q8 appears better at repeatedly checking its work, identifying problems and fixing them.</span></p><p><span style="font-size: 16px;"><br></span></p><h2>Impressive C++ Game Generation</h2><p><span style="font-size: 16px;">The model was also tested on a more difficult C++ skateboarding game. The Q4 version struggled to complete the task fully but still produced a surprisingly detailed result. It included elements such as a rider model and parts of a playable environment. The Q8 version went even further. Although it did not produce fully functional gameplay, it created NPCs, visual effects, camera controls and other elements. For a 35B model with only a small number of active parameters at a time, these results were considered particularly impressive.</span></p><p><span style="font-size: 16px;"><br></span></p><h2>3D Modeling Was Mixed</h2><p><span style="font-size: 16px;">The model was also asked to create a realistic six-cylinder engine that could theoretically be 3D printed.The Q4 version understood the basic concept and produced an engine-like model with six cylinders, but the overall detail was limited. The Q8 version created a much more detailed engine with additional components such as turbochargers and other engine parts. However, it had a major problem: the scale was badly incorrect. This means the Q8 result was more detailed, but the Q4 version arguably produced something more practical in terms of overall size.</span></p><p><span style="font-size: 16px;"><br></span></p><h2>Website Generation Was More Even</h2><p><span style="font-size: 16px;">Both versions were also asked to create a website featuring a 3D watch. The Q4 version produced a surprisingly acceptable result, including a watch, leather strap, reflections and several website elements. The Q8 version also created a good-looking result, with more realistic proportions in some areas. However, unlike the game-generation tests, there was no huge difference between Q4 and Q8 here. Both were capable of producing reasonable results.</span></p><p><span style="font-size: 16px;"><br></span></p><h2>The Q8 Advantage Was Clear in the Final Game Test</h2><p><span style="font-size: 16px;">The strongest difference between the two versions appeared in a new Street Yeet-style game test. The Q4 version managed to use Blender and Godot to create the game's assets and put together a working project. It had several problems and was not fully playable, but simply producing a functioning project was impressive.</span></p><p><span style="font-size: 16px;"><br></span></p><p><span style="font-size: 16px;">The Q8 version was considerably better. It created the assets, assembled the game and included working gameplay elements and effects. The result was much closer to what was requested and demonstrated the model's ability to work through a complicated multi-step task. Importantly, this particular test had not previously been shown to the model, making the result less likely to simply be the result of reproducing something from its training data.</span></p><p><span style="font-size: 16px;"><br></span></p><h2>Which Version Should You Use?</h2><p><span style="font-size: 16px;">For most people, the choice will come down to hardware. Q4_K_M is the better choice if you want something easier to run locally. It uses less memory and can still produce surprisingly capable results. For everyday coding, experimentation and simpler projects, it appears to offer very good value. Q8 is the better option if you have enough system memory to run it comfortably. The tests showed that it can provide substantially better results on complex coding, game development and agent-style tasks. The Q8 version can also potentially be partially offloaded into CPU RAM if there is enough available memory. The downside is that this can make the model slower.</span></p><p><span style="font-size: 16px;"><br></span></p><h2>Final Thoughts</h2><p><span style="font-size: 16px;">Ornith 1.5 35B-A3B is a surprisingly capable local AI model. The testing showed that it can do much more than basic text generation. It can build websites, create games, generate 3D models, work with development tools and autonomously troubleshoot some of its own mistakes. The Q4 version is particularly impressive for its hardware efficiency, while the Q8 version demonstrates noticeably stronger capabilities when memory isn't a major limitation. Neither version was perfect. Both produced broken features and required additional prompting or troubleshooting in some tests. However, the overall performance suggests that Ornith 1.5 35B could be an attractive option for people who want a powerful AI model that can run locally without needing extremely expensive hardware. For users interested in local AI, Ornith 1.5 35B is definitely a model worth trying, especially if you're looking for something between smaller consumer models and much larger frontier models.</span></p>