DeepSeek upgrades V4.1 Flash to V4.1-Flash-1017 with open MIT weights: Terminal-Bench 4.0 jumps from 31.2 to 42.3, same API price

Why it matters

【Key points】

On Oct 17 DeepSeek's official API changelog announced that DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017, with the web and app models switched over too. API usage and pricing are unchanged: deepseek-flash now serves the new model. Weights are on DeepSeek's official Hugging Face account under the MIT license. DeepSeek calls it a minor update to the Sept 10 V4.1 Flash, but says it improves substantially on reasoning, coding, agentic and visual benchmarks while cutting task runtime and output length.

Reported scores (old V4.1 Flash in brackets): Terminal-Bench 4.0 42.3 (31.2), ProgramBench 29.1 (20.3), HLE without tools 42.6 (36.8), Apex Bench v1.1 72.1 (64.6), Automation-Bench 60.5 (54.8), DeepSWE v1.1 77.0 (74.2). DeepSeek says it tuned the model for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and recommends DeepSeek Harness 0.2.1-rc.1 or later.

【Verifiable】

Official changelog: https://api-docs.deepseek.com/updates

Hugging Face: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017

【Our read】

1. Billed as minor, it gains about 11 points on Terminal-Bench 4.0 and about 9 on ProgramBench, mostly agentic and long-horizon coding, and it replaces the default model at the same price, so existing users get the upgrade automatically.

2. All numbers are self-reported, and the closed models in DeepSeek's table are Opus 5 and GPT-5.6 Sol, not the current Opus 5.5 or GPT-6 line, so this does not show parity with today's best closed models.

3. The model name briefly surfaced in DeepSeek's API list overnight and the community reads it as groundwork for a V4.1 Pro; there is no official word on that.