DeepSeek-V4-Flash Local Deployment Benchmarks: Performance Across Hardware

Following the release of DeepSeek-V4-Flash-0731, the community quickly initiated local deployment tests across various configurations, ranging from consumer-grade GPUs to DGX Spark. Performance varied significantly: a single RTX 3090 achieved only 4.02 tok/s in an unoptimized environment, while dual RTX PRO 6000 reached 243 tok/s using speculative decoding. Factors like quantization methods, VRAM capacity, memory bandwidth, and software stacks (such as vLLM and llama.cpp) all impacted the final results. These hands-on tests provide a hardware selection reference for users with different budgets and needs, demonstrating the feasibility of running open-source models on personal hardware.

已确认

多款消费级显卡可运行 DeepSeek-V4-Flash：RTX 3090（单卡 4.02 tok/s，双卡 12.5 tok/s）、双 RTX 3060（速度未明确）、3×RTX 3090（119GB 占用）、5 卡混插（PP 600 t/s）等。
AMD 显卡同样可行：3×MI50 达 15 t/s，混合 7900 XTX + 3×MI60 为 11 t/s。
DGX Spark 表现突出：单机 16.5 tok/s（3bit 量化），双机 82 tok/s（单流，峰值 95），并发 3 会话 135 tok/s。
投机解码显著提速：双 RTX PRO 6000 上从 243 tok/s（投机）对比无投机提升明显。
上下文长度受显存限制：双 RTX PRO 6000 最高 131K，但滑动窗口注意力可支持百万上下文（m16）。

尚未确认

部分性能数据来自单一测试，未复现，如 m14 的 243 tok/s 需进一步验证。
m9 中 DGX Spark 单机 agentic 工作流性能反超双卡 vLLM 的结论，依赖特定测试场景，普适性未知。

为什么重要

这些实测表明，DeepSeek-V4-Flash 在多种硬件上均可本地运行，性能足以满足个人或小团队使用，降低了前沿模型的门槛。同时，投机解码、量化等优化手段可大幅提升效率，为本地 AI 部署提供了实践参考。

2026-08-01 ~ 2026-08-03 · 21 related posts

Primary sources

Local LLM on Mac: M2 Ultra 192GB Long-Context Inference Benchmarks — Badger-Purple · 2026-08-02 23:43 (UTC+08:00)
DeepSeek-V4-Flash on RTX 3090 with 128GB RAM: 12.5 tok/s via --n-cpu-moe — Ok_Ninja7526 · 2026-08-02 05:22 (UTC+08:00)
DeepSeek-V4 Local Test: 17 t/s on A6000 + 256GB RAM — USBhost · 2026-08-02 11:16 (UTC+08:00)

Running DeepSeek V4 on a Single Unoptimized RTX 3090: 4 Tokens/sec — Altruistic_Heat_9531 · 2026-08-01
DeepSeek Hits 243 tok/s on Dual RTX 6000 with Speculative Decoding — TheZachMueller · 2026-08-02
DeepSeek v4 Flash Hits 82 tok/s on Dual DGX Sparks — andrewchen · 2026-08-02
Benchmarking DeepSeek on Dual RTX 3060 with 96GB RAM: Speed and Cost Analysis — esw123 · 2026-08-02
Multi-GPU Full VRAM Deployment of DeepSeek V4 Yields Only 600 t/s PP — fragment_me · 2026-08-02
Running DeepSeek V4-Flash Locally: Dual RTX 6000 Rig Handles Only Single User — dee_hw · 2026-08-02
DeepSeek V4 Flash 3-bit Quantization Tested on 3x RTX 3090: 119GB VRAM — consultkitapp · 2026-08-02
DeepSeek's New Release Significantly Boosts the Value of Nvidia DGX Spark — firstadopter · 2026-08-02
[source] DeepSeek-V4-Flash on RTX 3090 with 128GB RAM: 12.5 tok/s via --n-cpu-moe — Ok_Ninja7526 · 2026-08-02
DeepSeek-V4-Flash hits 44-59.5 tok/s on RTX PRO 6000 eGPU with llama.cpp — backslashHH · 2026-08-02
Custom IQ3 Requant for DeepSeek-V4-Flash: 1.4x Decode Speed on Mixed-GPU Rigs — HockeyDadNinja · 2026-08-02
Running DeepSeek V4 Locally on 3x MI50: Hits 15 t/s — Kamal965 · 2026-08-02
DeepSeek v4 Flash on Single DGX Spark Beats Dual-Spark Setup in Agentic Workflows — pbaylies · 2026-08-02
[source] DeepSeek-V4 Local Test: 17 t/s on A6000 + 256GB RAM — USBhost · 2026-08-02
Running Frontier Models Locally: DeepSeek V4 Flash on 2x DGX Sparks — lifebypixels · 2026-08-02
Running DeepSeek V4-Flash with 1M Context on 2x RTX PRO 6000: Concurrency Up 1.5x — dee_hw · 2026-08-02
Running LLMs on mixed AMD GPUs: 11 t/s inference speed achieved — Hyungsun · 2026-08-02
Running Agentic Coding on Two DGX Sparks: 100M Tokens/Day Economics — HarveenChadha · 2026-08-02
[source] Local LLM on Mac: M2 Ultra 192GB Long-Context Inference Benchmarks — Badger-Purple · 2026-08-02
Running DeepSeek V4 Flash 155G on DGX Spark: 2-bit Quantization & MTP Benchmarks — Puzzleheaded_Base302 · 2026-08-03
Running LLMs on DGX Spark: DeepSeek V4 Flash Tops the Ranking — andersonbcdefg · 2026-08-03