DeepSeek Swapped In V4.1-Flash-1017 Behind the Same deepseek-flash Id at the Same Price. Re-Run Your Evals
- What happened
- DeepSeek replaced V4.1 Flash with DeepSeek-V4.1-Flash-1017 behind the deepseek-flash name, with API calls and pricing unchanged.
- Why it matters
- Anything calling deepseek-flash already runs new weights that DeepSeek says produce shorter outputs, with no opt-out stated.
- What to do
- Re-run your evals and output-length checks on deepseek-flash, and move DeepSeek Harness to 0.2.1-rc.1 or later if you use it.
Our take: same name, same price, different model. If deepseek-flash sits in a production pipeline, you are already running 1017, so re-run your evals before you trust old results. DeepSeek's own numbers point up, and our DeepSeek V4.1 Flash verdict stays conditional until someone outside DeepSeek reproduces them.
Affected: DeepSeek V4.1 Flash (the model behind deepseek-flash), DeepSeek V4 Flash (its legacy id may now reach 1017 too) and DeepSeek Harness (DeepSeek recommends an upgrade).
What happened
DeepSeek's change log, 17 October: "DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version." It adds that "The API calling method and pricing remain unchanged" and that you "simply use deepseek-flash to access the latest model."
The entry mentions no opt-out and no pinned older snapshot.
DeepSeek's numbers: the 1017 column is from its 17 October entry, the previous column from its 10 September V4.1 Flash entry.
| Benchmark | Previous | 1017 |
|---|---|---|
| Terminal-Bench 4.0 | 31.2 | 42.3 |
| DeepSWE v1.1 | 74.2 | 77.0 |
| HLE (without tools) | 36.8 | 42.6 |
DeepSeek also says 1017 reduces task runtime and output length. Every number here is from DeepSeek's own runs. We have not tested 1017.
The Hugging Face card calls 1017 "a minor update to DeepSeek-V4.1-Flash", under the MIT licence, with a 1M-token context, and lists 763B parameters.
Why it matters
A model swap behind a stable id is a silent change to anything built on it. Prompts, parsers, eval baselines and cost estimates tuned on the old weights now run against new ones, and nothing in your logs will say so.
The shorter-output claim cuts both ways. Our conditional verdict on DeepSeek V4.1 Flash cites heavy verbosity, so terser answers would help on cost. They also break anything that expects the old output length or format.
Legacy ids are probably affected too. DeepSeek's 10 September entry says deepseek-v4-flash and deepseek-v4-flash-vision-exp are "temporarily routed to V4.1 Flash". Read together, the two entries suggest those names now reach 1017 as well. That is our inference: DeepSeek has not said it. If you need the frozen DeepSeek V4 Flash behaviour, self-host its weights.
What changes for you
- Re-run your evals on
deepseek-flash. The model behind it changed on 17 October, so baselines from before that date describe a different model. - Check output-length dependencies. Parsers, token budgets and cost forecasts built on the old output length need a fresh look.
- DeepSeek Harness users: DeepSeek recommends DeepSeek Harness 0.2.1-rc.1 or later "for the best experience" with 1017. Our directory rates DeepSeek Harness caution, so treat that as a version requirement, not an endorsement.
- Need reproducibility? Pin by self-hosting the weights you validated, since the API offers no stated way to stay on the old version.
FAQ
Did the price change? No. DeepSeek says the API calling method and pricing remain unchanged.
Can I stay on the previous V4.1 Flash through the API? DeepSeek's entry mentions no opt-out or pinned snapshot, so assume not.
Are the benchmark gains independent? No. All of them are DeepSeek's own runs.
What to do
- 1 Re-run your own evals on deepseek-flash: the model behind it changed on 17 October.
- 2 Check parsers, token budgets and cost forecasts that depend on output length, since DeepSeek says 1017 shortens outputs and task runtime.
- 3 If you use DeepSeek Harness, upgrade to 0.2.1-rc.1 or later, as DeepSeek recommends for 1017.
Affected tools & models
Never need to catch up again
The weekly delta: only verdict changes and act-now items. No digest filler.