RetroCraft - Qwen 3.8 27B Q8, one shot with exact performance data on dual 3090s.

The demo video above was created with a single prompt using Qwen 3.8 27B Q8_K_XL.

create me a single page, html/js/css GeoCities, Laser Wave, 80's retro, minecraft clone named RetroCraft. Embed it in a single go binary that serves the HTML page.

It took Opencode only a few minutes to write the code and it worked without an adjustments. There's going to be lots of posts and demos today but I wanted to show the actual pp, tok/s, tokens in/out, numbers for of the whole run.

llama-swap provides basic trajectory data and we can see that not every turn is has the same performance for prompt processing or generation:

ID Cached Prompt Generated Prompt Speed Gen Speed Duration Notes 251 43,473 95 90 142.46 t/s 48.76 t/s 2.76s final version 250 43,186 88 199 128.22 t/s 43.94 t/s 5.46s 249 42,757 255 173 311.94 t/s 40.11 t/s 5.37s prompt to improve cli flags 248 42,554 15 187 10.23 t/s 45.61 t/s 5.80s finished run 247 42,478 19 58 31.27 t/s 41.36 t/s 2.24s 246 41,330 194 953 91.89 t/s 45.97 t/s 23.07s 245 41,217 117 157 161.66 t/s 40.34 t/s 4.84s 244 41,007 19 190 32.43 t/s 65.54 t/s 3.71s 243 40,720 107 180 168.56 t/s 73.83 t/s 3.29s 242 40,032 521 168 464.05 t/s 77.38 t/s 3.50s 241 39,375 562 96 233.86 t/s 59.97 t/s 4.22s 240 39,194 185 510 256.77 t/s 39.97 t/s 13.69s 239 38,927 37 229 63.91 t/s 49.08 t/s 5.45s 238 38,719 28 179 49.58 t/s 44.10 t/s 4.82s 237 38,185 159 375 229.21 t/s 42.34 t/s 9.75s 236 37,818 55 311 89.87 t/s 52.56 t/s 6.72s 235 37,525 186 106 257.00 t/s 39.58 t/s 3.59s 234 37,317 20 187 33.91 t/s 62.15 t/s 3.79s 233 37,221 19 77 33.86 t/s 49.58 t/s 2.30s 232 36,801 19 400 33.45 t/s 55.57 t/s 7.94s 231 28,668 188 7,944 294.62 t/s 71.03 t/s 112.68s 230 28,342 118 209 198.16 t/s 51.91 t/s 4.77s start of agent implementation 229 8,931 51 19,361 70.58 t/s 40.89 t/s 475.21s initial prompt and reasoning. It did A LOT of thinking... Here are my llama-swap settings. The run was done using Q3.8-27B-MTP:preserve-thinking-coding in OpenCode. I found preserving the reasoning tokens greatly improves the results.

macros: "server-latest": | /path/to/llama-server/llama-server-latest --host 0.0.0.0 --port ${PORT} -ngl 999 -ngld 999 --no-mmap --no-warmup --log-verbosity 4 --fit off "qwen-27B-server": | ${server-latest} --temp 1.0 --min-p 0.0 --top-k 20 --top-p 0.95 --repeat_penalty 1.0 --presence_penalty 0.0 models: Q3.8-27B-MTP: sendLoadingState: false env: - "CUDA_VISIBLE_DEVICES=GPU-6f0,GPU-f10" filters: stripParams: "temperature, top_k, top_p, repeat_penalty, min_p, presence_penalty" setParamsByID: "${MODEL_ID}:thinking-coding": temperature: 0.6 presence_penalty: 0.0 "${MODEL_ID}:preserve-thinking-coding": temperature: 0.6 presence_penalty: 0.0 chat_template_kwargs: preserve_thinking: true enable_thinking: true "${MODEL_ID}:instruct": chat_template_kwargs: enable_thinking: false temperature: 0.7 top_p: 0.8 presence_penalty: 1.0 cmd: | ${qwen-27B-server} --model /path/to/models/Qwen3.8-27B-UD-Q8_K_XL.gguf # slightly too big for the full context --ctx-size 238000 --kv-unified --spec-type draft-mtp --spec-draft-n-max 3 --spec-draft-p-min 0.75 --tensor-split 50,50 -sm tensor -np 2 --batch-size 4096 --ubatch-size 512