I DID IT!! DeepSeek-V4-Flash-Vision EXL3 MixedK is now running VISION + DSpark speculative decoding together on ONE DGX Spark.
🚀🚀 19.7 tok/s text decode with DSpark3, ~2.2× my plain vision path, while image prompts still work correctly from the same vLLM process. 🔥 This is
One month in the arena.
Four models. $100,000 each. 44 assets, 2-hour windows, every trade on the permanent record since Aug 1. Kimi_Moonshot - Kimi K2 +3.96% claudeai - Fable 5.1 +0.10% Zai_org - GLM-5.3 Flash +0.01% deepseek_ai - DeepSeek V4 -1.53% Same market,
The one I showed at yesterday's study session
The one I showed at yesterday's study session: Using two DGX Spark (PGX) machines, I submitted 16 requests to DeepSeek-V4-Flash-Vision-Exp + DSpark at the same time. I used MiaAI_lab's recipe. The result is 62.6 tok/s for a single request, and a total of 248.8 tok/s for 16 requests (about 4 times more)!
[Lutheran News Agency | Dr. Yan said | Editing] Xi Jinping's "SCO Summit" series: empty words and nonsense, and small calculations making noise.
9/1/2026 DrLiMengYAN1 ◉ Xi Jinping now urgently needs a "safe environment" and is afraid of "external forces interfering in internal affairs"
You have reached the end of the archive
All of deepseek