DeepSeek Releases Model V4.1 Flash Update and Opens Serving Framework
In this article (3)
DeepSeek announced an updated V4.1 Flash artificial intelligence model alongside an open-source inference framework, pricing enterprise tasks down to roughly $0.20 each.
The developer posted a 77 per cent Pass@1 score on the DeepSWE v1.1 coding benchmark, cutting compute demands through shorter operational outputs.
Its release of V4.1-Flash-1017 on October 17 raised the system’s Terminal-Bench 4.0 score from 31.2 to 42.3, while DeepSWE v1.1 gained almost three points to 77.0. Benchmark evaluations show the system completing tasks for $0.11 to $0.14 when running lower-cost operational settings, scoring between 72 and 73 per cent. Independent tracking on Arena’s Agent Arena leaderboard placed the updated model first among open-weight competitors and 11th overall across 7,498 real-use sessions, with median execution costs between $0.18 and $0.19 per task.
Open Source Engine for Enterprise Chips
Three days after the model release, DeepSeek published DeepInfer, the inference engine running its production services, alongside a request-handling library named deepseek-recipe. The framework manages hardware allocation, latency schedules and computing kernels to serve queries at scale. It runs directly on Nvidia SM100 graphics processing units and Huawei Ascend 950 neural processing units, targeting V4.1 Flash deployments.
“Commercial access expanded in October with Search and Fetch APIs billed at 5 yuan per 1,000 successful queries.”
The open-source code does not provide an end-to-end service stack out of the box. Production teams still need to supply their own key-value cache infrastructure, load balancing across parallel replicas, and complete disaggregation systems. Developers gain internal scheduling kernels, but enterprise engineering teams will absorb the integration overhead.
Shifting the Cost of Business Automation
For online marketplaces, customer support platforms and software houses across Asia, the release attacks the cost ceiling that keeps autonomous agents out of routine commercial workflows. Running thousands of real-time inventory adjustments, code fixes and customer transactions through commercial western models remains too expensive for thin-margin Asian e-commerce. If an open-weight system matches proprietary grade accuracy at 20 cents a task, mid-sized digital retailers will self-host rather than pipe data through expensive closed APIs.
Hardware exclusivity remains the primary roadblock for regional operators seeking to replicate these results. While support for Huawei’s Ascend 950 gives Chinese domestic firms an alternative to export-controlled American silicon, Southeast Asian developers operating standard cloud setups will face hardware configuration bottlenecks. Running pruned agent frameworks demands deep systems engineering that many consumer brand operators do not maintain in-house.
Building Out the Infrastructure Layer
DeepSeek spent the previous quarter trimming resource requirements across its stack. The company cut key-value cache memory to a quarter and solid-state drive storage to an eighth of previous requirements during the September rollout of V4.1 Flash. It introduced off-peak API access at a 50 per cent discount to balance round-the-clock server loads.
Commercial access expanded in October with Search and Fetch APIs billed at 5 yuan per 1,000 successful queries. The firm is maintaining free API connections for developer platforms including Claude Code and Codex until per-request billing starts on November 1.