DeepSeek V4 Flash Sparks a Wave of Local Deployment Tests

DeepSeek-V4-Flash-0731 was rapidly tested across NVIDIA, AMD, Apple, and DGX Spark hardware after release. The broad takeaway is that the model can run on personal hardware, not just datacenter systems, but the gap between “it runs” and “it feels usable” is still large. In many cases, quantization choices, software backends, speculative decoding, and memory bandwidth matter as much as the underlying GPU.

Confirmed

• On NVIDIA, community results now span single consumer cards, pro cards, and multi-GPU setups. An unoptimized RTX 3090 setup reported 4.02 tok/s, while OkNinja7526 pushed a 24GB RTX 3090 with 128GB DDR5 to 12.5 tok/s on UD-IQ3S. A Reddit test relayed by USBhost showed about 17.2 tok/s on an RTX A6000 with 256GB DDR4 running Q8. Another benchmark showed UD-Q3KXL occupying 119GB on 3x RTX 3090. Higher-end setups went much further: backslashHH measured 44-59.5 tok/s on an RTX PRO 6000 Max-Q eGPU, and a separate developer reported a 243 tok/s median on 2x RTX PRO 6000 Blackwell using DSpark speculative decoding.
• Optimization clearly changes the picture. HockeyDadNinja said that re-quantizing only 129 router-expert tensors to IQ3XXS improved decode speed by 40% on mixed multi-GPU systems that spill into RAM. But raw hardware scale does not automatically solve usability: fragmentme loaded IQ3XXS fully into VRAM across five consumer GPUs and still saw only about 600 t/s prompt processing, concluding it was not yet ideal for everyday use.
• AMD, Apple, and DGX Spark all posted workable results as well. Kamal965 reached 15 tok/s with 3x MI50, and Hyungsun reached 11 tok/s with a 7900 XTX plus 3x MI60. On Apple hardware, milphickelson improved an M1 Ultra 128GB setup from 5-6 tok/s to 15-16 tok/s with a community engine patch, while antirez’s DwarfStar on a 512GB M3 Ultra reached about 37 tok/s on mxfp4. For DGX Spark, one test reported 16.5 tok/s for a 3-bit run, and another confirmed that a 2-bit build can fit the full 155GB model.

Unconfirmed

• A MiaAI result reposted by pbaylies claimed that a single DGX Spark reached 58.5 tok/s aggregate throughput at 12-way concurrency in an agentic workflow and outperformed a dual-GPU vLLM setup. Lucebox also claimed 51.1 tok/s on the full 284B model using heterogeneous consumer hardware with AMD. Both are notable, but this cluster does not include independent reproduction details.

Why it matters

These posts show that the local deployment barrier for a frontier MoE model is falling quickly across consumer PCs, Apple silicon, AMD systems, and compact AI machines like DGX Spark. They also underline that fitting the model is only step one: prompt throughput, unified or system memory bandwidth, and software-stack tuning remain decisive for real-world usability.

2026-08-01 ~ 2026-08-03 · 26 related posts

Full story(15 episodes)→

• Episode 1: DeepSeek-V4-Flash Architecture Leaked with Million-Token Context(2026-07-31, 3 posts)
• Episode 2: Unsloth Releases Quantized DeepSeek V4 Flash 0731 for Local Deployment(2026-07-31, 6 posts)
• Episode 3: DeepSeek on Huawei Ascend Beats OpenAI in Inference Profitability(2026-08-01, 2 posts)
• Episode 4: DeepSeek-V4-Flash Excels in Frontend Coding with Unmatched Cost-Performance(2026-08-01, 3 posts)
• Episode 5: DeepSeek V4 Preview: Flash to Introduce Four-Level Reasoning Effort(2026-08-01, 2 posts)
• Episode 6: DeepSeek Drastically Reduces Training Compute Costs Across Models(2026-08-01, 2 posts)
• Episode 7: DeepSeek V4-Flash API Public Beta Launches with Major Agent Upgrades(2026-08-01, 5 posts)
• Episode 8: DeepSeek V4-Flash Costs 105x Less, But Stability and Benchmark Overfitting Questioned(2026-08-01, 17 posts)
• Episode 9: DeepSeek V4 Flash Sparks a Wave of Local Deployment Tests(2026-08-01, 26 posts)
• Episode 10: DeepSeek V4 Flash Quantization Tests Show Major Speed Gains Without Quality Loss(2026-08-04, 2 posts)
• Episode 11: Single RTX 5090 Runs DeepSeek 1M Context(2026-08-04, 2 posts)
• Episode 12: DeepSeek-V4-Flash Tops Cost-Efficiency with Ultra-Low Running Costs(2026-08-05, 4 posts)
• Episode 13: DeepSeek V4 Flash Benchmarks Leak, Outperforming Pro(2026-08-05, 3 posts)
• Episode 14: DeepSeek V4 Flash Tops ARC-AGI Cost-Performance, Costing a Quarter of GPT-5.6(2026-08-08, 9 posts)
• Episode 15: DeepSeek V4 Flash Passes 22 Coding Tests on Dual DGX Spark Cluster(2026-08-10, 2 posts)

• 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
• [source] 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
• 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
• Local LLM on Mac: M2 Ultra 192GB Long-Context Inference Benchmarks — Badger-Purple · 2026-08-02
• [source] Running DeepSeek V4 Flash 155G on DGX Spark: 2-bit Quantization & MTP Benchmarks — Puzzleheaded_Base302 · 2026-08-03
• DwarfStar runs DeepSeek V4 Flash on M3 Ultra at 37 t/s — mishig25 · 2026-08-03
• M1 Ultra 128GB Test: Patch Boosts Local DeepSeek V4 to 16 tok/s — mil_phickelson · 2026-08-03
• Running LLMs on DGX Spark: DeepSeek V4 Flash Tops the Ranking — andersonbcdefg · 2026-08-03
• DeepSeek-V4-Flash hits 48.9 tok/s on single DGX Spark with speculative decoding — max_paperclips · 2026-08-03
• Lucebox Beats NVIDIA DGX Spark by 3.6x in DeepSeek Decode — samsja19 · 2026-08-03
• Run DeepSeek V4-Flash with 1M Context Locally on Dual RTX PRO 6000 — dee_hw · 2026-08-03