# DeepSeek-V4.1-Flash-1017: a same-size update with better agent scores and shorter answers

> OnPrem World · https://onpremworld.com/en/news/2026-10-deepseek-v4-1-flash-1017 · CC BY 4.0

Published: 2026-10-17 · Models

DeepSeek released V4.1-Flash-1017 on October 17, a minor update to V4.1-Flash with open weights on Hugging Face and ModelScope under the MIT licence; it is also live on DeepSeek's website, app and API. DeepSeek says it improves reasoning, coding, agentic and visual results while finishing tasks faster with shorter outputs. In DeepSeek's own table it scores 42.3 on Terminal-Bench 4.0 against 31.2 for V4.1-Flash, and 77.0 on DeepSWE v1.1 against 74.2. The checkpoint is the same size as V4.1-Flash on Hugging Face, and it again ships without a Jinja chat template.

## Why it matters locally

Anyone already running V4.1-Flash locally can swap in the new weights with the same memory budget, while for everyone else the model is still far too big for a single workstation.

## For the market

DeepSeek keeps refreshing an MIT-licensed model in place instead of waiting for a new version number, and its own table puts the update ahead of GPT-5.6 Sol on Terminal-Bench 4.0 (42.3 versus 39.9) while still trailing Opus 5 (51.8). Shorter outputs cut cost per task for API users, which likely matters more to hosted-AI buyers than to people running it on their own servers.

## In plain words

DeepSeek has released an improved version of its large free AI model, better at finishing multi-step jobs such as writing a whole program, and it gets there with fewer words. It is about the same size as before, so it still needs a very large computer to run at home or in an office.

## For practitioners

Hugging Face lists the same parameter total as V4.1-Flash, so the sizing from our September item still applies: start with a 512 GB machine such as an M5 Ultra Mac Studio or a multi-GPU server, and confirm your runtime loads the new checkpoint before replacing the old one. There is still no Jinja chat template, so use DeepSeek's encoding.py reference or the deepseek-recipe library, and start from the recommended settings of temperature 1.0, top_p 1.0 and max_tokens of at least 256K.

## Source

[DeepSeek-V4.1-Flash-1017 release note (DeepSeek API Docs)](https://api-docs.deepseek.com/news/news261017)

## On this site

- Models: [DeepSeek-V4-Flash](https://onpremworld.com/en/models/deepseek-v4-flash.md)
- Hardware: [M5 Ultra](https://onpremworld.com/en/hardware/apple-m5-ultra.md)

[All news](https://onpremworld.com/en/news.md)

Summaries are written by OnPrem World from the official source. We link to the original and never republish it.