The KING Is BACK! DeepSeek V4.1 Flash 1017 Runs Almost 2x Faster on 4x DGX Spark

DeepSeek V4.1 Flash just got its 1017 update, and on my 4x NVIDIA DGX Spark cluster it's now running almost TWICE as fast. I switched the engine from vLLM to TensorFold, a native inference engine, and the local speeds went crazy. In today's video we break down what's new in DeepSeek V4.1 Flash 1017, how it stacks up against Claude Opus 5 and GPT 5.6 Sol on DeepSeek's own benchmark table, where it lands on the Artificial Analysis index, and the full speed results running it 100% locally on four DGX Sparks. ⚡ The numbers (4x DGX Spark, TP4, EXL3 3.5 bpw, 1M context): • 1 user: 93.6 tok/s on TensorFold vs 48.5 on vLLM • 6 users: 213.8 tok/s total (vLLM: 165.7) • 93K-token prompt read in 42.1 seconds (the first V4.1 Flash took 78.2) • KV cache pool: 2.1M tokens (+38%) • Update: the latest speed sheet hits 271 tok/s total at 16 users 🆕 UPDATE: WE DID IT ON 2 SPARKS! DeepSeek V4.1 Flash 1017 now runs on just 2x DGX Spark (TensorFold TP2 + LibertAI EXL3 2.9 bpw): • 1 user: 73 tok/s · 4 users: 136 tok/s total • Full 1M context (found a needle in a 1,000,000-token prompt) • 93K-token prompt read in 43.7 seconds • Tool eval 97.1, the same score as the 4-Spark setup 2-Spark recipe: https://github.com/tonyd2wild/DeepSee... Quick note: the extra speed boost came from a second network LANE on the same cable, not a second cable. Each Spark has one cable, and TensorFold now uses both lanes inside it. ⏱️ Chapters 0:00 DeepSeek V4.1 Flash 1017 just dropped 0:46 DeepSeek's release timeline 1:42 Benchmarks vs Claude Opus 5 and GPT 5.6 Sol 3:25 Artificial Analysis index 4:13 vLLM vs TensorFold on 4x DGX Spark 5:02 Speed results: almost 2x faster 5:13 KV cache pool and the second network lane 5:43 93K-token prompt: 78 seconds down to 42 6:05 LibertAI EXL3 quants and model sizes 6:53 Will it fit on 2 DGX Sparks? 8:26 TensorFold speeds by task 9:13 GLM 5.3 Flash vs DeepSeek V4.1 Flash 9:57 Build with cloud, run local 🔗 Links • Full recipe (scripts, patches, benchmarks): https://github.com/tonyd2wild/DeepSee... • DeepSeek V4.1 Flash 1017: https://huggingface.co/deepseek-ai/De... • LibertAI EXL3 3.5 bpw quant: https://huggingface.co/LibertAIDAI/De... • TensorFold engine: https://github.com/ashhart/TensorFold 🙏 Credits: TensorFold (ashhart), bertholomus for the DeepSeek V4.1 TP4 fork and recipe, LibertAI for the EXL3 quants, and DeepSeek for the model. Let me know in the comments: would you run DeepSeek V4.1 Flash locally, or stick with the cloud? Subscribe for more local AI, NVIDIA DGX Spark, GPU builds, AI models, and tech discussions! #DeepSeek #DGXSpark #LocalAI #TensorFold #Tech2Wild ━━━━━━━━━━━━━━━━━━ 💻 GitHub → https://github.com/tonyd2wild 🐦 Follow on X → https://x.com/Tech2Wild ━━━━━━━━━━━━━━━━━━