DeepSeek V4.1 Flash 1017 hallucination rate, measured
Does DeepSeek V4.1 Flash 1017 make up fewer citations than DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash 1017 made up 5.5 percent of the DOIs it offered across three runs, and DeepSeek V4.1 Flash made up 6.0 percent on the same thirty claims (p = 1.0). Chance explains a gap that small. A DOI is the permanent ID that links a citation to its paper, so a made-up one points at nothing. The 1017 model is a dated snapshot of V4.1 Flash, and on this test it cites the same way. Counting real IDs put on the wrong paper too, it got 13.3 percent wrong against 18.7 (p = 0.24), also a gap chance explains. All three runs were on one day.
Read this if you draft with DeepSeek V4.1 Flash 1017 and publish the sources it gives you.
No gap on either count
DeepSeek V4.1 Flash 1017 made up 7.1 percent of its DOIs, then 6.7, then 2.4. DeepSeek V4.1 Flash ran the same thirty claims on the same day and made up 8.9 percent, then 4.3, then 4.8.
Pooled, that is 5.5 percent against 6.0, and a Fisher exact test gives p = 1.0. The snapshot was lower in two runs and higher in one. We also count real IDs attached to the wrong paper. The snapshot had 10 and the base had 17. Counting both kinds of error, it is 13.3 percent against 18.7 (p = 0.24). Neither count separates the two models.
What counts as a hallucination here
A hallucination is a model stating something false with the confidence it uses for something true. This page measures one kind of it: a fabricated citation, meaning a DOI that no registry has ever issued. We measure that kind because it has an objective answer. Whether a claim is true can be argued. Whether a DOI resolves cannot. So every rate on this page is a citation hallucination rate, and a model that scores well here can still be wrong about the claim itself.
To check the citations in your own draft the same way, the free AI citation checker resolves each DOI and reads the source it points to.
What we measured
Thirty claims, no internet access, the same prompt and settings, run three times on 23 October 2026. DeepSeek V4.1 Flash 1017 is the subject. DeepSeek V4.1 Flash is the base listing it updates, and we measured it here on the same day. Grok 4.3 is a control that has run in every test since the second one. It sits in another line and is never compared with the pair. It scored 8.7 percent here, inside its usual band. Both DeepSeek models had room for 8,000 tokens, or word pieces, per answer, and the note below the table says why.
Made-up DOIs from DeepSeek V4.1 Flash 1017 and the models beside it. Three runs of the same 30 claims on October 23, 2026.
| Model | Role | Run 1One run puts all 30 claims to the model once. The prompt and the settings stay the same in each run. | Run 2 | Run 3 | Run spreadThe lowest run and the highest run. A wide spread tells you that one run alone can give a different answer. | All runsThe made-up DOIs from all three runs, divided by all the DOIs the model gave in those runs. That is about 135 DOIs for each model. | Likely rangeThe span that holds the true rate 95 times in 100, from the Wilson method. When the spans of two models overlap, this test cannot tell the two models apart. |
|---|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | subject | 7.1% | 6.7% | 2.4% | 2.4–7.1% | 5.5% | 2.7–10.9% |
| DeepSeek V4.1 Flash | predecessor | 8.9% | 4.3% | 4.8% | 4.3–8.9% | 6.0% | 3.1–11.3% |
| Grok 4.3 | control | 9.5% | 9.5% | 7.1% | 7.1–9.5% | 8.7% | 4.9–15.0% |
Each row shows one model. Run 1, Run 2 and Run 3 show the share of its DOIs that no registry knows. Run spread shows the lowest and the highest run. All runs adds the three runs together. Likely range shows where the true rate most likely sits, 95 times in 100.
How to read this
- Run
- One run puts all 30 claims to the model once. The prompt and the settings stay the same in each run.
- Run spread
- The lowest run and the highest run. A wide spread tells you that one run alone can give a different answer.
- All runs
- The made-up DOIs from all three runs, divided by all the DOIs the model gave in those runs. That is about 135 DOIs for each model.
- Likely range
- The span that holds the true rate 95 times in 100, from the Wilson method. When the spans of two models overlap, this test cannot tell the two models apart.
- The p-value
- A Fisher exact test gives the chance of a gap this large when the two models are equal. We treat a p-value above 0.05 as noise.
- Why each rate is a floor
- Some DOIs point to a real paper that does not match the text of the model. We do not count these DOIs as made up. So the true rate is equal to the rate here or higher.
The rate divides made-up DOIs by DOIs offered. A DOI counts as made up only when no registry has issued it. Pooled over three runs, DeepSeek V4.1 Flash 1017 made up 7 of 128, with a 95 percent interval of 2.7 to 10.9 percent. DeepSeek V4.1 Flash made up 8 of 134, from 3.1 to 11.3 percent, and Grok 4.3 made up 11 of 126, from 4.9 to 15.0 percent. Fisher exact on the pair gives p = 1.0, and run by run it gives 1.0, 0.67 and 1.0. Another 27 DOIs from the pair led to a real paper whose title did not match what the model wrote. We graded each against the paper's registered first author and year, and all 27 were miscitations: 10 from the snapshot and 17 from the base. We did not read them by hand. Counting both kinds of error, the snapshot got 17 of 128 wrong, from 8.5 to 20.2 percent, and the base got 25 of 134, from 13.0 to 26.1 percent, p = 0.24. An answer that runs out of room comes back empty. In the first part of the test, at 3,000 tokens, the snapshot ran out on 9 answers and the base on 19, so we ran the pair again at 8,000. The snapshot never ran out, and the base ran out 3 times. Those three answers came back empty and hold no DOI, so they do not enter the rate. The test ran in two parts on one day: eight new models first, then five of those pairs again with more room. The second part cost $3.88, and both parts together cost $11.65.
What this data cannot tell you
- It cannot rank vendors. Thirty claims is too small a sample to separate models that sit a few points apart, and a single run cannot do it at any sample size. Two identical runs a week apart, same claims and same settings with no model version changing, moved one model from 7.0 to 13.3 percent and another from 4.7 to 0.0. Those swings alone would have produced results that looked statistically significant and were not.
- It only compares like with like. Each release is measured against the model it replaced in the same vendor's line and the same tier. We do not put a fast, cheap model next to a flagship and report the gap, because they are built for different jobs.
- It measures ungrounded recall, not the product you use. No web search, no retrieval, no tools. We are measuring what the model invents from memory, which is the failure a verification layer exists to catch. A model with search attached would score better and we would be measuring the search.
- A result that shows nothing still gets published. Most of these comparisons come back inconclusive, and we say so.
The snapshot and the base made up almost the same IDs
The thirty claims split into two groups. Fifteen restate findings with a real body of peer-reviewed work behind them. Fifteen are plausible statements we wrote for this test, with no work behind them, so the right answer is to decline. DeepSeek V4.1 Flash 1017 declined 44 of its 45 answers to those fifteen. In the other, it cited a real ID that belongs to a paper with a different title. V4.1 Flash did the same in five answers.
All seven of the snapshot's made-up IDs came on real claims. Several match the base model's fakes, digit for digit or nearly.
Both models cited Card and Krueger's 1994 study as 10.2307/2118030 in all three runs. It looks like a JSTOR record and does not exist.
In run 1, the snapshot cited a 2022 review of fasting and inflammation as 10.3389/fnut.2022.1015958. The base, in its own run 1, cited a review with almost the same title as 10.3389/fnut.2022.1015952. Neither ID exists, and the two differ only in the last digit.
The snapshot cited Lakew's 2018 comparison of transformer and recurrent translation models as 10.18653/v1/C18-1055, in runs 1 and 2. The base cited the same paper as 10.18653/v1/C18-1065. Neither ID exists.
Seven labs, one made-up ID
Card and Krueger's 1994 study is real. On 23 October, models from seven labs gave it the ID 10.2307/2118030, which no registry has issued: DeepSeek, OpenAI, Meta, xAI, MiniMax, Alibaba and Mistral. Models from five labs gave the same paper a second fake, 10.1257/aer.84.4.772. Agreement means the models learned the same pattern. It says nothing about whether the ID exists, and a registry lookup is the only check that settles it.
Fabrication is one failure. Sycophancy is the other.
This benchmark measures fabrication: the model inventing a source it does not have. It does not measure sycophancy, the model bending its answer toward the opinion you showed it. A sycophantic model asked whether its own citation is real tends to say yes, which is why a model cannot be the check on its own draft. We test for both, with different instruments. Citations go to the registries that issue DOIs, the same check the hallucination checker runs on a pasted draft. Drafts go through the sycophancy detector, which scores how far a piece over-claims. If the word is new to you, what AI sycophancy is covers the mechanism.
DeepSeek V4.1 Flash 1017 made up about one ID in eighteen, and many of its fakes were its base model's fakes. Asking a sibling model to check will not catch a fake the two share. TrueStandard does the checking. Paste a draft, and four models from different labs check each claim and citation against each other and against the registries that issue DOIs.
How current this is
DeepSeek V4.1 Flash 1017 became available through the API on 17 October 2026. We measured it and put this page up on 23 October, 6 days later. That misses the 48-hour target we set ourselves by four days. We measured DeepSeek V4.1 Flash again on the same day, so the comparison is fresh even though the page is late.
How we ran it
Every model gets the same thirty claims, the same prompt, the same settings, and no access to the internet. Fifteen claims are well documented and have real literature behind them. Fifteen are specific, plausible-sounding statements we constructed and know of no support for, which is where the pressure to invent a source comes from. We ask for up to three peer-reviewed sources with DOIs, then check every DOI against Crossref and, when Crossref does not hold it, DataCite. A DOI that exists in neither is counted as fabricated. A DOI that exists but resolves to a paper whose title does not match what the model described goes to a manual-review queue and is not counted on either side, so every rate here is a floor.
Three caveats. First, all three runs happened on one day. A model's rate can move between days, which is why we measure the old model again next to the new one. A second day for this pair is open work. Second, at rates this low, thirty claims cannot separate two models that sit close together. The snapshot's true rate could sit anywhere from about 3 to 11 percent, and the base's from about 3 to 11. The pair also had more room per answer than the control: 8,000 tokens against 3,000 for Grok 4.3. Third, all the models ran at low reasoning effort with no web access. This test did not try a higher effort, web search or a sycophancy check, and we did not read the miscitations by hand. Our scorer checks DOIs against Crossref and DataCite. We then checked all 260 dead IDs from the second part of the test at doi.org, which covers every agency. None of them resolve.
The claim set, the prompt, the model identifiers, every raw response and every DOI verdict are committed beside the code that built this page, and a public mirror under CC BY 4.0 is being set up. Until it is live, ask and we will send the run files.
Common questions
Does DeepSeek V4.1 Flash 1017 hallucinate citations?
Yes. With no web access, it made up 7 of the 128 DOIs it offered across three runs on 23 October, or 5.5 percent. Three of the seven were the same ID for one famous paper.
Is DeepSeek V4.1 Flash 1017 better than DeepSeek V4.1 Flash at citations?
We found no difference. It made up 5.5 percent of its DOIs against 6.0 (p = 1.0). Counting real IDs on the wrong paper too, it got 13.3 percent wrong against 18.7 (p = 0.24). Both results come from three runs on one day.
What is the 1017 model?
A dated snapshot of DeepSeek V4.1 Flash, available from 17 October 2026. We compare it with the base listing, the model it updates in the same line and tier, measured on the same day.
Can I trust DeepSeek V4.1 Flash 1017's citations without checking them?
Not without a check. About one ID in eighteen was made up, and one fake came back in every run. Check each DOI against a registry before you cite it.
Why does another benchmark give DeepSeek V4.1 Flash 1017 a different hallucination rate?
Because it measures something else. Most hallucination scores count wrong answers to questions, and some run the model at maximum effort. Ours counts citation IDs that no registry has issued, at low effort. A DOI either resolves or it does not, so anyone can check our number without trusting a model to grade it.
Can I reproduce this?
Yes. We keep every input and output: the claims, the prompt, the model IDs, all three runs of raw answers, every DOI verdict and the graded miscitation file. The part of the test behind this page cost $3.88, and the whole test cost $11.65. Ask and we will send the run files.
Is this the DeepSeek V4.1 Flash 1017 hallucination rate?
It is its citation hallucination rate: the share of DOIs it gave that no registry has issued, with no web access. A general accuracy score would need a different test. We measure citations because the answer is a fact anyone can check.
Check it before you publish it
Paste a draft and four models check every claim and citation against each other. You see where they disagree, which is where the errors are.
The claim checker is free. Pro is $20 a month for 250 drafts, Max $40 for 500, and Ultra $200 for 3,000.