DeepSeek V4.1 Flash 1017 在单台 128GB DGX Spark 上跑通：39.4 tok/s
约 114GB VQ 量化在单台 DGX Spark 上实测：解码 39.4 tok/s、32K 预填约 1145 tok/s，跑通 110K 提示；Artificial Analysis 智能分从 39 升到 45，每任务成本 $0.49 降到 $0.26。
So a few days ago I said DeepSeek V4.1 Flash 1017 had been squeezed from a ~510GB checkpoint → ~114GB, specifically targeting a single 128GB DGX Spark… …but nobody had actually benchmarked it on the Spark yet. Now they have. 👀 One DGX Spark: 🧠 DeepSeek V4.1 Flash 1017 💾 ~114GB VQ build 🖥️ 111–113GB memory used 💿 203GB Engram tables streamed from SSD ⚡ 39.4 tps weighted decode 🚀 ~1,145 tps prefill @ ~32K 📚 tested with a 110K-token prompt So yeah, a single 128GB Spark is actually serving it. Meanwhile, the independent scores are in. Artificial Analysis: Previous V4.1 Flash → 39 New 1017 → 45 🚀 That's +6 points and puts it equal #2 among open-weight models on their Intelligence Index. AA also measured: 💰 cost/task → $0.49 → $0.26 ⚠️ The Spark build is an aggressive ~1.6-bit VQ, so this is not full-checkpoint quality.
这条是引用着另一条写的：x.com/teksedge/status/2111547801620394454