Three days after release, DeepSeek's V4.1-Flash-1017 has turned the open-source charts into its home turf: No.1 on both of Arena's core open-source boards and No.11 overall — at roughly one-tenth the cost of peer models. The phrase "Pareto improvement of performance vs. cost" is now trending across the AI community.

A glowing neural network with a golden light column symbolizing chart-topping open models
A glowing neural network with a golden light column symbolizing chart-topping open models

The surprise drop on October 17

On the evening of October 17, DeepSeek shipped V4.1-Flash-1017 with weights on HuggingFace and ModelScope. Benchmarks rose "across the board" versus the previous Flash, with clear gains in coding and agent tasks; DeepSWE tests show its performance-cost frontier pushed further out across DeepSeek Harness, Tencent CodeBuddy and other agent frameworks. Official wording: a Pareto improvement — better work, fewer tokens.

Key facts

· No.1 on both Arena open-source boards, No.11 overall (Oct 19–20)

· Top open-source model on the WebDev board, with costs cut ~90% vs. rivals

· New Search + Reader APIs: per-1K-successful-request pricing, ~half a cent per call, failures free

The showpiece is "wish-list coding": state what you want, and the model plans, writes, runs, checks and fixes on its own — official demos include a browsable 3D Forbidden City, a web racing game and a Godot remake of Portal 2. The companion self-developed Search API fills the agent-browsing gap, with site, language and time filters.

Blue code blocks interwoven with golden data streams
Blue code blocks interwoven with golden data streams

Beyond the charts: pricing power for open weights

The release itself was telegraphed by leaks — a briefly exposed model-list entry and a changed system fingerprint. But after the double-chart win, community talk has shifted from scores to deployment cost: as the gap between open models and commercial flagships narrows while the price keeps a zero, the economics of self-hosted inference stacks are tilting fast.

When performance converges upward and price converges toward cost, open models become the industry's pricing anchor.

For NineZenith, high-value open models like this are a key foundation option for the TianXing model matrix: every order-of-magnitude drop in inference cost accelerates large-scale agent deployment. We track and fold the best model-cost combinations into enterprise delivery.

(Compiled from public reports by QbitAI, Sina Finance, Geekpark and AIBase)