DeepSeek's October 17 snapshot is a shorter, cheaper company chart
Vercel lists DeepSeek-V4.1-Flash-1017 as a minor gateway update with shorter runs. The Decoder reports DeepSeek's own chart, including a cheaper DeepSWE score. Not ours.
The short version
The Decoder says DeepSeek shipped V4.1-Flash-1017 on its site, app, and API, with weights on Hugging Face and ModelScope, and that the company's own chart shows higher scores than the previous snapshot at a lower cost. Vercel lists the snapshot on AI Gateway as deepseek/deepseek-v4.1-flash-1017 and calls it a minor update that shortens agent runs. The scores belong to DeepSeek's chart, as The Decoder reports it.
A snapshot with a company chart
The Decoder's October 17, 2026 account says DeepSeek released V4.1-Flash-1017 on the company's website, app, and API, and published the weights on Hugging Face and ModelScope. It says the company calls the update a Pareto improvement: better scores than the previous V4.1 Flash snapshot, finished sooner, with shorter outputs. Vercel's changelog the same day says the snapshot is on AI Gateway under the id deepseek/deepseek-v4.1-flash-1017, describes it as a minor update, and says agent runs tend to take less time and use fewer output tokens because outputs are shorter. The Hugging Face URL is the card The Decoder points at. This brief does not add parameters or mixture-of-experts counts from a card it is not quoting. The scores below are the company's, through The Decoder. The artificial-intelligence desk isArtificial Intelligence.
The DeepSWE line, attributed
The Decoder says a cost chart for DeepSWE shows 77.0 percent at an average just under $0.20 a task, and a low setting still above 72 percent at about $0.10. It says the predecessor scored 74.2 percent at about $0.50 a run. A chart the vendor drew is evidence of what the vendor claims. It becomes a public measurement when someone else can run it. A page on this paper about what a second lab has to be able to rerun is the standard this number does not yet meet.an evaluation a second lab can run.
Shorter output is the cost mechanism Vercel names
Vercel does not print the DeepSWE table. It does say the snapshot is more concise than the previous one, and that concision is why runs use less time and fewer output tokens. The Decoder's dollar figures are one way to price that concision. Vercel's sentence is the qualitative version. They agree on direction, shorter, and they do not agree on a shared table because only one of them printed a table. The paper isFutureweb.
The search API is a second announcement
The Decoder says DeepSeek is also opening Search and Fetch APIs, priced at $0.75 per 1,000 calls, with free access through some harnesses until October 31, 2026, at 4:00 p.m. UTC. It says the company describes the backend as its own search engine. Vercel's gateway note does not set that price. This brief repeats the price as The Decoder's account and does not pretend the model card has been checked for the same sentence. A free window with an end time is a schedule, not a permanent term. Software isSoftware.
Rival names stay inside the chart
The Decoder lists larger models DeepSeek's chart claims to beat in some coding and agent categories. Naming them here would spread a vendor comparison into this paper's voice. The fact this page will keep is that The Decoder says the chart makes those comparisons. A reader who wants the names can read The Decoder. A reader who wants an independent run does not have one on this page. Hardware isHardware.
A gateway id is a distribution fact
deepseek/deepseek-v4.1-flash-1017 is a string Vercel says its gateway will accept. Availability on a gateway is not a quality result. It is a second place a caller can reach a snapshot the company also offers on its own API, according to The Decoder. Two doors, one model name, still one vendor chart. Security is Asked directly, Where does Vercel say the model can be called? Vercel says the gateway id is deepseek/deepseek-v4.1-flash-1017. It calls the snapshot a minor update to DeepSeek V4.1 Flash that improves reasoning, coding, agent tasks, and image understanding, and it says runs tend to take less time and use fewer output tokens.Security.
What this page will not treat as measured
This page will not say the snapshot is cheaper in a workload this desk ran. It will not say weights were audited. It will say DeepSeek shipped a dated snapshot, Vercel is serving that id, The Decoder reports a company chart with the DeepSWE figures above, and a search API price appears in The Decoder's account with an October 31 free window. Open source, for a survey of what organizations say they use, isthe 2027 survey.
Who is in scope
Callers who can select the gateway id, and readers of the company's chart, are in scope. People who need a reproducible eval are in scope only as a warning: the number is not yet that eval. A harness version The Decoder mentions is a compatibility note for the company's own tools, not a requirement this page is issuing. Networking is Asked directly, What DeepSWE numbers does The Decoder attribute to DeepSeek? The Decoder says the company's chart shows 77.0 percent at just under $0.20 a task, and that a low setting still clears 72 percent at about $0.10, against a predecessor at 74.2 percent and about $0.50. Those are DeepSeek's figures, as The Decoder read the chart.Networking.
Where the snapshot sits
The Decoder is the chart and the API price. Vercel is the gateway listing and the shorter-output sentence. The Hugging Face URL is the weights page The Decoder names. Nothing here is this paper's benchmark. The short answer this page will stand on is this: The Decoder says DeepSeek shipped V4.1-Flash-1017 on its site, app, and API, with weights on Hugging Face and ModelScope, and that the company's own chart shows higher scores than the previous snapshot at a lower cost. Vercel lists the snapshot on AI Gateway as deepseek/deepseek-v4.1-flash-1017 and calls it a minor update that shortens agent runs. The scores belong to DeepSeek's chart, as The Decoder reports it.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
Where does Vercel say the model can be called?
Vercel says the gateway id is deepseek/deepseek-v4.1-flash-1017. It calls the snapshot a minor update to DeepSeek V4.1 Flash that improves reasoning, coding, agent tasks, and image understanding, and it says runs tend to take less time and use fewer output tokens.
What DeepSWE numbers does The Decoder attribute to DeepSeek?
The Decoder says the company's chart shows 77.0 percent at just under $0.20 a task, and that a low setting still clears 72 percent at about $0.10, against a predecessor at 74.2 percent and about $0.50. Those are DeepSeek's figures, as The Decoder read the chart.
Does this page confirm the card text on Hugging Face?
The Decoder says the weights are on that Hugging Face page and on ModelScope. This brief names the page The Decoder names. It does not add architecture numbers from a card this desk has not quoted line by line.
What does The Decoder say about a search API?
It says DeepSeek is opening Search and Fetch APIs at $0.75 per 1,000 calls, and that some harnesses can use them free until October 31, 2026, at 4:00 p.m. UTC. That price is The Decoder's account of the company's schedule. It is not a Vercel sentence.
Are rival-model wins this paper's measurement?
The Decoder says DeepSeek's chart beats named larger models in several coding and agent categories. A second lab that can run the same evaluation is a different standard. This page does not adopt the wins as its own measurements.