DeepSeek V4.1 Flash 1017: 763B Model Cuts Costs 60%, Boosts Coding & 3D

On October 17, DeepSeek upgraded its V4.1 Flash model to version V4.1-Flash-1017 across web, app, and API endpoints without requiring any API call changes. The accompanying DeepSeek Harness framework was also updated to 0.2.1-rc.1.

Model Scale and Open Weights

The new model contains 763 billion parameters, up from V4 Flash (284B) and V4.1 Flash (552B). DeepSeek's Flash series now exceeds the parameter count of its earlier flagship models V3 and R1. For comparison, GLM-5.3 stands at 744B. Model weights are openly available under the MIT license on Hugging Face ( https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) and ModelScope (

https://modelscope.cn/models/deepseek-ai/DeepSeek-V4.1-Flash-1017

).

Benchmark Improvements

Official evaluations show significant gains over the September 10 V4.1 Flash release:

Terminal-Bench 4.0: 31.2 → 42.3 (+35.6%)

ProgramBench: 20.3 → 29.1 (+43.3%)

DeepSWE v1.1 pass rate: 74.2% → 77.0%

Cost Efficiency

The max-tier average cost per task dropped from approximately ¥3.3 to ¥1.3 (a 60% reduction) while achieving higher pass rates. Inference speed remains around 250 tokens/second. DeepSeek attributes the savings to improved agent task execution efficiency: the model shortens runtime and output length, and is optimized for agent frameworks like DeepSeek Harness, CodeBuddy, and WorkBuddy. API pricing per token remains unchanged.

"Wish-Style Programming" Capabilities

The highlight is the ability to build complete projects from scratch with minimal prompts and few interaction rounds. Using self-feedback reinforcement learning, the model enhances game development and 3D modeling. Demonstrated examples include:

3D modeling of the Forbidden City

A browser-based racing game

A Godot-engine replica of Portal 2

A dynamic 3D scene visualizing prehistoric to modern history

Early community tests confirm the improved performance.

DeepSeek Search API Launch

DeepSeek also opened its proprietary search engine API. The engine indexes hundreds of billions of web pages globally and handles hundreds of millions of daily search requests. Two endpoints are provided:

Search – query web pages

Fetch – retrieve full page content

Both are billed per successful request at ¥5 per 1,000 calls (¥0.005 each). Failed, empty, or timed-out requests are not charged. Starting November 1 (Beijing time), server-side search/fetch calls made via DeepSeek Harness, Claude Code, or Codex will incur this fee.

Usage Recommendations

API users continue to call deepseek-flash to access the new model. DeepSeek recommends upgrading Harness to 0.2.1-rc.1 or later for optimal compatibility.