Post Content
[[{“value”:”DeepSeek V4 Flash (Official) Runs Locally: Benchmarks, Cost, and DGX Spark Setup
I test the official DeepSeek V4 Flash (July 31 edition) running entirely locally and break down DeepSeek’s released benchmarks, where it beats the prior V4 Pro/Flash preview and even outperforms larger models like GLM 5.2, while approaching Opus 4.8 for a fraction of the size. I also explain why benchmark harnesses matter (including DeepSeek’s agent harness and an OpenAI harness example) and highlight internal DS Bench results. I cover pricing (about $0.02/M input tokens and ~$0.30/M output tokens), token efficiency on the AI Index cost curve, and agentic coding results like Terminal Bench. Then I demo Open Code harness runs (Pokémon encyclopedia and an International Space Center tracker), show ~20–30 tok/s on two DGX Sparks in FP4, discuss the 1M context window, hardware/VRAM needs for 4-bit and 3-bit, and options like Dwarf Star for SSD offloading and mixed precision.
Thanks to @NVIDIADeveloper for DGX Sparks.
LINK:
https://www.deepseek.com/en/
https://api-docs.deepseek.com/guides/responses_api/
https://github.com/antirez/ds4
https://x.com/UnslothAI/status/2083231049434435596
https://artificialanalysis.ai/#intelligenceDwarfStar
https://x.com/stevibe/status/2083120066678464750
Video: https://youtu.be/9gHcmhUDJfw
Long Running Agents: https://youtu.be/ju7R6jer6_M
My voice to text App: whryte.com
Website: https://engineerprompt.ai/
RAG Beyond Basics Course:
https://prompt-s-site.thinkific.com/courses/rag
Signup for Newsletter, localgpt:
https://tally.so/r/3y9bb0
0:00 – Introduction to DeepSeek V4 Flash
0:55 – Shocking Benchmark Results & Performance Efficiency
0:2:42 – The Impact of Optimized Harnesses on Performance
0:4:23 – Pricing & The “Pareto Frontier” of Cost vs. Intelligence
0:5:04 – Running DeepSeek V4 Flash Locally
0:7:20 – Live Demo: Agentic Coding & Pokemon Encyclopedia Test
0:8:34 – Live Demo: ISS Tracker with Real-time UI
0:9:39 – Limitations: Text-only & Multimodal Alternatives
0:10:51 – Hardware Requirements for Local Setup
0:12:52 – Final Results & Summary”}]] Read More Prompt Engineering
#Promptengineering #AI