 | One way I like to test new models, is by one-shoting (with a good prompt) a single webpage clone of the classic arcade game pacman. I usually do 3 attempts and keep the best one. So far all of them, including anthropic, chatgpt and google models, have failed, most of them miserably. The best one until now was GLM 5.1 That was until I tried it with Qwen 3.6 27b F16. Out of 3 attempts, 2 were the best by far, with the top result only having minor errors! However, as soon as I dropped to 8bit quantisation, I could not replicate those good results even after trying 5+ times. This goes to show what I have saying for a long time, based on my experience: there is a world of difference between a 16bit and a 8bit quant, despite most people claiming it is lossless, or nearly lossless. The results were so good, and since it just happened that I was testing the llama.cpp MTP speculative decoding PR (not yet merged at that time) with my own quants, and developing my own fixed jinja chat template for Qwen 3.5/3.6, I thought why not try to push Qwen 3.6 27b F16 through a proper agentic coding workflow. I think the results were brilliant, and they speak for themselves. You can try the full single page game here: Lessons learned and observations: * A good chat template is critical. The official chat template was unusable due to it being only targeted at vLLM, and therefore full of errors in other tools. I started with community templates, which were improvements, but still had many quirks. This is why I started fixing the bugs one by one in the official templates, and slowly improving it. The beginning of the agentic sessions were painful due to many quirks and errors. But slowly it improved, and once I got the template well tuned, it felt like I had unlocked a new level of intelligence in the model. * MTP speculative decoding does not accelerate all tasks identically. Basically it is most efficient at deterministic task like coding, and least at creative tasks like brainstorming. I wrote about it here: https://www.reddit.com/r/LocalLLaMA/comments/1t9gcar/mtp_benchmark_results_the_nature_of_the/ - For this pacman development, my generative tok/s varied between 8 tok/s and 18 tok/s depending on the task. For reference, without MTP, I get 6.6 tok/s with the same model and quant. * Not all harnesses are equals both in terms of code quality but also in terms of impact on speed. Most of use already know that the coding harness has a huge impact on quality, with Claude Code being considered the gold standard; this is what I use for normal daily coding. In this case I started with Qwen CLI, mostly because of the chat template problems, on the principle that if there was one harness more likely to better handle Qwen LLM specifics, it would be their own harness. I was actually pleasantly surprised, and Qwen CLI delivered far beyond what I was expecting! In the later stages, I switched back to Claude Code, mostly to verify that the final chat template was working properly there too. I did not notice any improved process or code quality. What I noticed though, is that developing in Claude Code was a lot slower than in Qwen CLI! This is due to all the extra prompts built within Claude Code. With a local model that has such a slow tok/s, it can make the difference between being usable, and between being borderline hair pulling... * Context management and caching is super efficient in this model. Do not interfere with it. It works great, let it do its thing. Do not use any skill, plugin, etc, that manipulates the cache or context. This will result in confusing the model and making it a lot dumber and error prone. * Tool calls, context compaction, shell usage, subagents, parallel subagents, work flawlessly. Initially it did not though, and it took me a long time and lots of work to get it right through chat template fixes and improvements. I actually only used context compaction for testing, and it was fine, as usual in Claude Code. * High context is usable without too much degradation. Maximum context size is 256k tokens I believe. Most of the time I planned the tasks to stay below 100k, but there were a few times I pushed it slightly over 150k. I did notice slightly reduced capabilities, but nothing major. The main reasons why I tried to keep it low is to get the best reasoning capabilities, as with all other models, but also speed started to decrease as the context usage grew. * Apart from Gemini, this is the first model that impressed me with its audio knowledge. As a composer, musician, psychoacoustic scientist, and audio engineer, I pay a lot of attention to good audio. In this case, I tasked it to do some advanced audio manipulation and creation. All the audio in the game comes from Qwen having programmed the web audio synthesizer in a highly advanced and complex way. This is not midi, not simple wavetables, not samples. It takes into account psychoacoustic properties tuned to human hearing, with the use of harmonics, distorsion, layers, various effects. Truly impressive work. The only exception is the waka-waka sound, for which I had to make it use a sample (the same method was used in the original arcade game). * I can live with slow token generation speed. I used to think that I needed a minimum of 70 to 80 tok/s for viable development. But this was usable, gave me time to do other things in parallel, and also to better reflect on the agentic tasks. I would probably not use it for large projects, with my current hardware, but for small to medium project, it is definitely acceptable. If you read until here, let me know what you think, and I hope you enjoy the game. Dev environment: macOS, apple silicon M2 max, 96GB RAM, llama.cpp server with OpenAI and Anthropic API endpoints. [link] [comments] |