Skip to content
Startups

DeepSeek Releases Model V4.1 Flash Update and Opens Serving Framework

By Sarah Chen
2 min read
deepseek
deepseek
In this article (8)

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.

Questions & Answers

Q.

How does the updated model reduce operational costs for e-commerce businesses?

A.

The V4.1 Flash model lowers enterprise task costs to roughly $0.20 each, with lower-cost settings completing tasks for $0.11 to $0.14. This pricing allows digital retailers to automate routine inventory updates and customer transactions without relying on expensive proprietary Western programming interfaces.

Q.

What hardware is required to run the DeepInfer framework?

A.

DeepInfer runs directly on Nvidia SM100 graphics processing units and Huawei Ascend 950 neural processing units. While this provides Chinese firms with a domestic alternative to export-controlled silicon, regional operators using standard cloud configurations may face hardware configuration bottlenecks.

Q.

What technical components must engineering teams build themselves to deploy the framework?

A.

The open-source release does not provide a complete end-to-end service stack. Engineering teams must supply their own key-value cache infrastructure, load balancing across parallel replicas, and disaggregation systems to deploy the model in production.

Q.

How is DeepSeek pricing commercial API access following the update?

A.

DeepSeek charges 5 yuan per 1,000 successful queries for its Search and Fetch application programming interfaces, alongside a 50 per cent discount for off-peak access. Free connections for developer platforms remain available until per-request billing begins on 1 November.

Reader pulse

Will 20-cent open-weight AI replace proprietary western models in Asian retail?

17,337 votes so far

Background from the Retail Academy

Background guides from the Retail Academy, free to read.

  • E-commerce

    Omnichannel fulfilment

    Ship-from-store, click and collect, marketplace fulfilment and live commerce, what each model demands from inventory accuracy, staffing and systems, and wher…

    Retail Operations · 11 min read

  • E-commerce

    E-commerce in Brunei

    Which platforms matter in Brunei, how buyers pay, why cash on delivery still shapes the economics, and what a delivered order really costs.

    Asia Market Guides · 9 min read

  • General

    AI in retail

    A practical map of where AI already pays off in retail, from demand forecasting and pricing to search, customer service and store operations, and where the h…

    Retail Operations · 11 min read

  • General

    Loyalty and membership design

    Points, tiers or paid membership: which mechanic fits which business, what a loyalty programme actually costs, and how to measure incremental effect instead…

    Retail Leadership · 11 min read

Weekly Briefing

Asia's retail intelligence, in your inbox

Monday, Wednesday and a Friday Weekly Wrap: the retail stories, numbers and moves that mattered across Asia. Nothing else, and you can unsubscribe in one click.

  • Top industry moves and market shifts
  • Weekly data-driven analysis from across Asia
  • Monday, Wednesday and the Friday Weekly Wrap

Read by retail operators, investors and brand teams across Asia.

Protected by a quick human check. No spam, ever. Unsubscribe in one click.

SecureGDPR ready