单张 RTX 5090 跑 V4.1 Flash 建体素世界：74 分钟耗尽上下文
Maya-S 量化 340GB 单卡运行 748B 模型，体素世界任务 74.1 分钟写完 10 万 token，浏览器检查 8.5/10，代码解码 22.9 tok/s，两次都因 131K 上下文将满而止。
DeepSeek-V4.1-Flash-1017 on one RTX 5090 ran out of context room at minute 74.1 of a voxel-world build, on @PeasantSmith's Project Maya. On the 5090 I normally only run the steampunk airship dock prompt. A model this big ran so well here that I gave it the voxel world too. His Maya-S quant, 340 GB on disk, one slot. A 748B model on a single card, most experts read from RAM and the NVMe SSD. Agent: DeepSeek Harness 0.2.1-rc.2. One prompt, one try, no help. Thinking on, sampling from the DeepSeek model card: temperature 1.0, top_p 1.0. Voxel world: done in 74.1 min Tokens written: 100,575 Requests: 21 Prompt of the last request: 121,159 tokens Last answer: 16 tokens, then the harness stopped Browser check: 8.5/10 Decode, code: 22.9 tok/s (separate speed run, thinking off) Decode, prose: 18.9 tok/s 500 words: about 35.1 s Eight requests at once: 19.7 tok/s in total The day before, same box and settings, other prompt (the airship dock): Airship dock: done in 74.6 min Tokens written: 98,217 Requests: 32 Prompt of the last request: 125,831 tokens Browser check: 9/10 Open: the context was nearly full both days (121,159 and 125,831 of 131,072 tokens). Both runs ended with a last answer of 16 tokens, because my test proxy keeps a safety margin. The last saved page counts, and it loads without errors. The model card allows answers up to 262,144 tokens. My test proxy trimmed the answer limit of 14 of the 21 requests to what still fits in 131,072. 2 of the 5 saved versions had a page error on the way. The final one loads clean. Measured on Project Maya v1.1.22. Newer releases have come out since. Every number, the exact command and the raw output: https://niklaslenz.com/v2/deepseek-v41-flash-5090-maya-s-mayads4/ Project Maya and the Maya-S quant by @PeasantSmith. DeepSeek-V4.1-Flash-1017 by @deepseek_ai.
这条是引用着另一条写的：「单张 RTX 5090 跑 748B 的 V4.1 Flash：Maya-S 量化实测」