DeepSeek's latest release has earned a new distinction. The DeepSeek-V4.1-Flash-20261017 model is now the highest-ranked open model on Arena.ai's Code Arena WebDev leaderboard, overtaking Kimi K3 (Max) while running at roughly one-tenth of the cost.

What the Leaderboard Shows

Arena.ai posted the updated rankings on October 18, 2026. The new DeepSeek build scored 1670 points on Code Arena WebDev, placing it #1 among open models and #12 overall. The leaderboard covers agentic and front-end coding tasks.

Kimi K3 (Max), the model displaced at the top of the open rankings, is estimated to cost approximately 10 times more to run, according to Arena.ai. The platform also framed the release as shifting the performance-cost Pareto frontier for open models in workflows — in other words, the trade-off between what a model delivers and what it costs to run.

A Point Release With Outsized Gains

The October 17 build is technically a minor update, refreshing the DeepSeek model that launched in September 2026. Minor, in this case, still means better on every benchmark DeepSeek listed. Comparisons against the September version include:

• DeepSWE v1.1: 77.0, up from 74.2
• Terminal-Bench 4.0: 42.3, up from 31.2
• Automation-Bench: 60.5, up from 54.8

Terminal-Bench 4.0 posted the largest jump of the three, rising 11.1 points.

Under the Hood

The model is built on a Mixture-of-Experts design, an architecture in which only a subset of parameters is activated for any given task. The backbone holds 552 billion parameters, with an additional 196 billion Engram parameters on top, but the model does not fire all of them at once. Because inference compute scales with active rather than total parameters, this sparse setup keeps the work behind each generated token in line with a far smaller network.

During prefill, the stage in which the model reads and processes input, 8 billion parameters are active. During decoding, when it generates output, that figure scales to 16 billion.

Other listed features include a Causal Encoder-Decoder structure, native vision (the model can take images as direct input), and KV cache optimizations that reduce the memory needed to store attention states during generation. The context window stretches to 1 million tokens, long enough to hold substantial codebases or extended agent sessions in a single context.

The weights are published on Hugging Face under the MIT license, a permissive license that allows commercial use, modification, and redistribution — the practical difference between a model teams can run and modify on their own infrastructure and one they can only access through an API. For developers who would rather not self-host, DeepSeek's API is priced at $0.30 per million input tokens and $1.20 per million output tokens at peak.

Background: DeepSeek's Playbook

DeepSeek, formally Hangzhou DeepSeek, is a Chinese AI lab that has built its identity around open-weight models. Its V4.1 family leans specifically into cost-efficient agentic coding.

What It Means for Developers

The most direct impact lands on engineering teams choosing which model to wire into their tooling. A model that ranks #12 overall, against both open and closed competitors, while charging $0.30 per million input tokens presents a cost-competitive option for coding workloads.

The ranking also means 11 models still sit above it. The very top of the leaderboard remains out of reach for now. The next Arena.ai leaderboard update will show whether the October 17 build holds the open-model lead it just took.