{"id":1988,"date":"2026-10-05T08:27:40","date_gmt":"2026-10-05T08:27:40","guid":{"rendered":"https:\/\/blog.openzeka.com\/en\/?p=1988"},"modified":"2026-10-05T08:35:11","modified_gmt":"2026-10-05T08:35:11","slug":"qwen3-8-flash-next-what-a-single-dgx-spark-or-rtx-pro-6000-can-do","status":"publish","type":"post","link":"https:\/\/blog.openzeka.com\/en\/qwen3-8-flash-next-what-a-single-dgx-spark-or-rtx-pro-6000-can-do\/","title":{"rendered":"Qwen3.8-Flash-Next: What a Single DGX Spark or RTX PRO 6000 Can Do"},"content":{"rendered":"

Qwen3.8-Flash-Next is a 176B-parameter Mixture-of-Experts model with a 262K-token context window. It scores 39.8 on the Artificial Analysis Intelligence Index<\/strong> and 53.6 on the Agentic Index<\/strong>.* Its NVIDIA NVFP4 checkpoint is 132.7 GB. A DGX Spark has 128 GB of memory, and an RTX PRO 6000 has 96 GB, so on paper the model fits on neither.<\/p>\n

In our measurements it runs on both: at 28.5 tok\/s on one DGX Spark<\/strong> and 155.8 tok\/s on one RTX PRO 6000<\/strong>. This post explains how, and what each device delivers.<\/p>\n

The trick: a table the model looks up, not computes with<\/h2>\n

To generate each token, a language model runs through all its layers and multiplies by its weight matrices. Those weights, and the KV cache that remembers the conversation, are read at every step. They must sit in fast GPU memory, or everything slows down.<\/p>\n

Qwen3.8-Flash-Next also carries a different kind of component: a 51B-parameter lookup table<\/strong> of learned word combinations, known as an Engram-style table. The model does not compute with this table; it looks things up<\/strong> in it. For each token, it reads only the handful of rows that match the last few words: about 2.5 KB out of a 51 GB table. DeepSeek, which introduced the idea, calls this conditional memory<\/em>: a row is read only when the input calls for it.<\/p>\n

Think of it as a large dictionary on a shelf. You do not need to carry the whole book around; you open only the page you need.<\/p>\n

So the table does not need to be in GPU memory. It can sit in a slower, larger place while the GPU keeps the part of the model it computes with:<\/p>\n