DeepSeek V4.1 Flash 1017: MIT weights, a silent API swap

If your app calls deepseek-flash, it is running a different model than it was yesterday. On October 17, DeepSeek released DeepSeek-V4.1-Flash-1017, published the weights under the MIT license, and pointed its Flash API at the new model. Same model ID, same call. Different brain.

What actually happened

From DeepSeek's announcement, its Hugging Face model card and its API docs:

• Open weights, MIT license. The license covers the repo and the weights. Hugging Face lists 763B total parameters and a 1M-token context window. DeepSeek publishes no GPU or memory requirements, and no hosted inference provider serves it on Hugging Face yet.
• The API swap. DeepSeek's docs say deepseek-flash "is now served by DeepSeek-V4.1-Flash-1017." The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but they route to the new model and bill at the Flash price.
• The claimed gains. DeepSeek says the update scores higher on agent benchmarks and writes shorter outputs. Its own numbers: Terminal-Bench 4.0 rises from 31.2 to 42.3, and DeepSWE v1.1 from 74.2 to 77.0. These are vendor-reported, with no independent runs yet.
• Search and Fetch APIs open to everyone. $0.75 per 1,000 calls per endpoint. Calls through DeepSeek Harness, Claude Code and Codex server-side tools are free for now. Billing starts October 31 at 16:00 UTC.

Why it matters for your business

Shorter outputs and better agent scores sound good. But "better on a benchmark" and "the same on your prompts" are not the same thing. A JSON extractor that worked yesterday can drop a field today. A support bot tuned for long answers now answers short. You did not ship a change, and your output still changed.

We have written before about DeepSeek retiring model aliases. This is the same lesson with a sharper edge: a hosted model ID is a pointer the vendor moves, not a version you own.

What we would do this week:

• Rerun your eval set against deepseek-flash now. Compare against last week's outputs. If you have no eval set, this is the week to build 30 real examples.
• Watch output length. Shorter answers can break parsers, truncation logic and cost forecasts in both directions.
• Decide if the weights change your plan. MIT weights mean you can pin this exact version on your own hardware or a host you pick. At 763B parameters, that is a serious GPU bill, so price it before you commit.
• Put a date on the search API. If an agent uses DeepSeek search, it gets a bill on October 31.

Sources: DeepSeek announcement, Hugging Face model card, DeepSeek API docs.