Back to Project
D

DeepSeek Flash Local Evaluation

Project Oct 18, 2026

Infatoshi's community measurement bundle for a locally served, attention-edited DeepSeek V4.1 Flash 1017 checkpoint on four Blackwell workstation GPUs.

Visit primary link 2 sources attached
Infatoshi's aggregate output-throughput chart shows about 132, 425 and 830 tokens per second at concurrency one, eight and 32 on four Blackwell GPUs.
Author-published interim local measurements on four shared RTX PRO 6000 Blackwell GPUs. Output includes reasoning; the points are leg means and the bars show observed min-max, not confidence intervals or single-user decode rates. Full chart preserved; not independently reproduced by AITrending. Source: Infatoshi
self-hostedcoding-agentsverification

What happened lately

DeepSeek Flash Local Evaluation timeline

Infatoshi published this evaluation bundle on October 18, 2026, with structured timing results, capability and behavior measurements, pinned model-edit provenance, charts and reproduction code. It measures a native attention-edited DeepSeek V4.1 Flash 1017 reference on four RTX PRO 6000 Blackwell Workstation Edition GPUs while other workloads share the hardware.

The bundle is interim research evidence, not a model-weight release, an official DeepSeek score or a finished comparison of quantized runtimes. Its new benchmark code and results carry MIT terms; upstream models, datasets and runtime components retain their own licenses. AITrending has reviewed the published artifacts without reproducing the GPU runs.