DeepSeek makes V4.1-Flash-1017 faster and cheaper than DeepSeek-V4.1-Flash, hitting 42.3 on Terminal-Bench

Top Model · 1,057 Likes

If you're choosing a model for agent runs, DeepSeek positions V4.1-Flash-1017 as faster and less costly to use than DeepSeek-V4.1-Flash. This minor update also produces shorter outputs and improves reasoning, coding and visual benchmark results.

Here's what changed:

• Terminal-Bench 4.0 rises to 42.3 from 31.2 for DS-V4.1-Flash.
• DeepSeek reports 77.0 on DeepSWE v1.1, versus 74.2 for DS-V4.1-Flash and 74.0 for Sol Opus 5.
• Automation-Bench v1.0.6 reaches 60.5, up from 54.8 for DS-V4.1-Flash.
• The MIT-licensed model and weights support local inference, with a 1M-token context window.

One catch: There is no Jinja-format chat template. The provided deepseek-recipe toolkit handles request encoding and output parsing, but leaves inference, tool execution and HTTP transport to you.

Try it: Use deepseek-ai/DeepSeek-V4.1-Flash-1017 and the repository's inference folder for weight conversion and local inference. Also available via Vercel

Sources: Hugging Face