Developing more powerful artificial intelligence models is only half the task; the second half involves ensuring their availability for large-scale operation. DeepSeek announced the V4.1 Flash update, aimed at accelerating task execution and shortening output lengths, and released an open-source framework for a more efficient inference process.
On October 17, DeepSeek released the V4.1-Flash-1017 model, which is an update to its V4.1 Flash model. According to the company's published test results, improvements were achieved in several areas, including coding tasks and agentic tasks. For instance, the Terminal-Bench 4.0 score increased from 31.2 to 42.3, and DeepSWE v1.1 rose from 74.2 to 77.0.
The cost-performance ratio diagram for the DeepSWE v1.1 coding benchmark shows that V4.1 Flash-1017 achieves a Pass@1 rate of 77% at a cost of about $0.20 per task, surpassing several models that require significantly higher operating costs. At lower settings, ranging from $0.11 to $0.14 per task, the model still demonstrates results around 72–73%.
DeepSeek also reports reduced task execution time and smaller output volumes, which helps lower inference costs because the model performs the same work using fewer tokens and less computational time. This update is optimized for agentic frameworks such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.
However, it should be noted that the presented figures are primarily the company's own test results, and different tests use different evaluation settings and frameworks, which does not guarantee the model's performance across all applications.
A separate signal came from the Agent Arena leaderboard on Arena. According to an ITHome report from October 20, the updated model ranked first among open-weight models and 11th overall, based on 7498 real-use sessions. The median cost was approximately $0.18 per task, although values ranging from $0.18 to $0.19 appeared on different pages of the leaderboard.
On October 20, DeepSeek introduced DeepInfer, its internal inference framework, as well as the deepseek-recipe library for handling API requests, prompt preparation for the model, and result parsing. Inference frameworks define how models utilize computational resources to process requests. Their design affects latency, throughput, and the hardware required to serve users, making them an important, albeit less visible, factor in AI operational costs.
DeepSeek describes DeepInfer as a framework that enables its large-scale production services to operate. The open-source release includes optimized compute kernels, asynchronous scheduling, and support for distributed inference and speculative decoding, according to the README and deployment documentation. It currently supports Nvidia SM100 GPUs and Huawei Ascend 950 NPUs, with DeepSeek V4.1 Flash listed as a supported model.
This release does not constitute a complete inference stack. DeepSeek's documentation indicates that a full service requires a KV-cache system, prefill/decoding disaggregation, and load balancing between replicas with data parallelism, which are not fully covered by this release, although most of the core disaggregation functionality has already been implemented.
Providing the code as open source gives developers access to part of DeepSeek's model serving technology. Practical utility will depend on their hardware, workloads, and the additional engineering effort required for deployment.
When launching V4.1 Flash in September, DeepSeek reported reducing KV-cache memory requirements by a quarter and SSD storage needs by an eighth compared to the previous generation. The company also offers off-peak API rates at half the price of peak rates.
In October, DeepSeek opened its API Search and Fetch to developers, setting the price at 5 yuan per 1000 successful requests. Calls through DeepSeek API integrations with Claude Code and Codex are temporarily free, with billing for requests scheduled for November 1.
These achievements demonstrate that DeepSeek is working to improve efficiency at both the model and infrastructure levels while expanding the toolkit for AI agents. Whether these improvements translate into reliable performance and reduced overall costs will depend on real-world workloads, deployment, and maintenance.
