AITOPAI热报
ModelsMulti-source

DeepSeek V4 Flash 家用部署实测：单机峰值 1875 tok/s

Geek@geekbb
Post· 9 sources
View the post
Why we picked it

有人晒了 DeepSeek V4 Flash 在 DGX Station 上的实测数据，单机每秒能跑 1875 个 token，128k 上下文首字只要 5 秒，你可以拿这个数去衡量本地部署值不值。

Alec Fong 在单台 GB300 DGX Station 上部署 DeepSeek V4 Flash，测得聚合峰值吞吐 1875 tok/s，单流 300 tok/s，128k 上下文下首 token 延迟 5 秒。推文感叹能在家部署该模型的人相当于 80 年代万元户，令人羡慕。

Original · Geek

现在能在家部署 DeepSeek V4 Flash 的就相当于80年代的万元户，90年代有私家车，让人羡慕。 Alec Fong @alecqfong The DGX Station is a beast at serving DeepSeekV4 Flash On a single GB300 Station I’m getting Peak aggregate 1875 tok/sec Single stream 300 tok/sec TTFT 5s for 128k context 🔗 View Quoted Tweet📊 1 ⚡

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include DeepSeek V4 Flash, DGX Station, inference performance, local deployment.

Source attribution: curated from Geek (@geekbb). Cross-source corroboration: 9 independent sources have covered this development (including NVIDIA AI, pandaily, IT之家), confirming industry-wide signal strength.

#DeepSeek V4 Flash#DGX Station#inference performance#local deployment

Keep reading

Recent stories with the same tag or section
PapersBuildersMulti-sourceSelected00:02
SelfSearch coding agent optimizes its own harness on Terminal-Bench

Harness quality, not just model choice, drives benchmark scores — this shows a cheap automated search can close the gap with hand-engineered harnesses like Codex. Builders running agent evaluation pipelines can apply the self-modification loop to cut inference costs and lift pass rates on their own stacks. The claim that it was built on DeepSeek V4 Flash also signals open models can compete with frontier harnesses when the scaffolding is tuned.

#SelfSearch#DeepSeek V4 Flash#Terminal-Bench#SWE-bench
Event
elvis@omarsar0Source
PapersBuildersSelected09:54
SelfSearch: Reward-Free Search for Self-Improving Agents

SelfSearch demonstrates that experience-based self-improvement can match evaluation-guided search at significantly lower cost, making agent development more efficient. Developers working on autonomous systems can implement this approach to reduce computational expenses while maintaining performance, particularly valuable for resource-constrained environments or when iterating rapidly on complex agent architectures.

#SelfSearch#agents#Terminal-Bench#SWE-bench
Event
arXiv: DeepSeek@Jungwoo Yang et al. (3)Source
PapersBuildersSelected10:46
SelfSearch: Reward-Free Search for Self-Improving Agents

SelfSearch demonstrates that agents can improve through self-modification experience without expensive downstream evaluations, significantly reducing search costs while maintaining competitive performance. Developers working with autonomous systems can implement this approach to enhance their agents' capabilities more efficiently, particularly for tasks requiring frequent self-improvement like coding assistants or reasoning models.

#SelfSearch#agents#Terminal-Bench#DeepSeek V4 Flash
Event
arXiv: DeepSeek@Jungwoo Yang et al. (3)Source