Is LLM PDF Translation Too Expensive? Testing a New pdf2zh-next Workflow: Kimi for Terminology + DeepSeek for Body Text

Introduction

#

Anyone who has used pdf2zh-next knows that it performs very well on complex layouts, formulas, and cross-page tables. But it has always had one pain point: LLM translation consumes too many tokens and costs too much.

Powerful models are expensive, while cheap models often collapse in translation quality. Is there a way to save money without sacrificing quality?

There is. Recently, pdf2zh-next added a key feature: it supports separate configuration for “terminology extraction” and “body translation.” In simple terms, one model first reads the whole document, extracts proper nouns, and fixes their translations; then those terms are passed to the translation model for full-text translation. This prevents the same term from being translated inconsistently and reduces the risk of the model improvising.

To find the best combination, I tested seven documents from different fields with several mainstream Chinese LLMs: Kimi K2 (0905), DeepSeek V3, and DeepSeek V3.2 (Exp). The goal was to see which model is more reliable for terminology extraction.

1. Extraction Quantity: More Is Not Always Better

#

First, look at the number of extracted terms. I tested three models on seven different files, with the results below:

Table 1: Number of extracted terminology entries by model

At first glance, the DeepSeek series generally extracts more entries. But after checking the details, I found that many of the extra entries are not useful. DeepSeek tends to capture years, DOI links, and even formula numbers as “terms.” These invalid entries are not helpful and instead increase token consumption and unnecessary cost.

By contrast, Kimi K2 extracts fewer terms, but the signal-to-noise ratio is higher, with far fewer filler entries.

2. Quality Deep Dive: Which Model Understands Academic Text Better?

#

Quantity alone means little. Accuracy matters. I manually checked the extracted results and found that each model has its own personality. Here are several representative cases.

1. Kimi K2: Restrained and Rigorous

#

K2’s strongest impression is that it is very rule-abiding. When it is unsure about a proper noun, it would rather keep the English original than translate recklessly. In academic translation, this is an advantage.

Its clearest strength is that it uses established academic translations. In File 1, it translated ACM CSUR accurately as “ACM Computing Surveys,” matching the meaning of Computing Surveys. V3 and V3.2 both rendered it as “ACM Computing Investigation,” but anyone familiar with computer-science academia knows the former is the accepted term. Also in File 1, for Sapienza University of Rome, K2 gave the official Chinese name “Sapienza University of Rome.” V3.2 simply transliterated it as “Rome Sapienza University,” while V3 translated it as “Rome Wisdom University.”

In Files 3 and 6, K2 even recognized book and journal titles, adding Chinese book-title brackets to System of Logic and MIS Quarterly, translating them as “System of Logic” and “MIS Quarterly.” These details make the translation feel more academic.

Most importantly, K2 is safe. In File 2, there was a big trap: Enterprise Rent-A-Car. K2 honestly preserved the English. But both DeepSeek versions forcibly translated it as “Avis Rent A Car.” Avis is Enterprise’s competitor. If this error appeared in a formal context, the consequences could be serious.

K2 also has a weakness: it can be too conservative. In File 4, for O-enhancing HRM practices, K2 translated it literally as “O-enhancing HRM practices,” which reads stiffly. V3.2 helpfully expanded it as “opportunity-enhancing HRM practices.” In the same file, for multiple NCA, K2 only translated it as “multiple NCA,” unlike the V3 series, which expanded it to “multiple necessary condition analysis.”

2. DeepSeek V3.2: Smart and Fluent, but Prone to Guessing

#

V3.2 produces the smoothest Chinese. It is good at converting stiff terms into readable Chinese, which helps reading. But over-interpretation also leads to factual errors.

Its strength is that it can make terms sound natural. In File 2, Evangelistic customer loyalty became “evangelistic customer loyalty” in V3.2, which captures the idea of customers actively recommending something like evangelists. K2’s “gospel-style customer loyalty” is literal but stiff; V3 used “fanatical customer loyalty.” In File 1, V3.2 also recognized the obscure abbreviation AIBOM as “AI bill of materials,” while K2 and V3 kept “AIBOM.”

Its weakness is clear: it over-paraphrases and sometimes hallucinates.

The most typical example is visible colleges in File 6. V3.2 may have associated it with the better-known invisible college and translated it in the opposite direction as “invisible colleges,” reversing the meaning. V3 translated it as “visible colleges,” while K2 did not extract the term.

In File 7, V3.2 was also too aggressive with personal names. The text included A. Ng, and V3.2 uniformly translated all instances as “Andrew Ng.” Andrew Ng is indeed A. Ng, but not every A. Ng in a paper should automatically be treated as him. K2’s decision to keep “A. Ng” is safer.

For Amazon Mechanical Turk in File 7, K2 preserved the English, while V3.2 and V3 both translated it as “Amazon Turkish robot,” misunderstanding the crowdsourcing platform as a robot.

3. DeepSeek V3: Somewhere in the Middle

#

V3 sits between the other two. Its semantic stability is slightly better than V3.2 because it is less eager to rewrite concepts, but its problem is higher noise.

It first produces too much invalid information. In the terminology table for File 6, V3 captured many year ranges, such as “1980–2000” and “1937–2016,” and full DOI URLs such as “https://doi.org/10.1016/…”. These do not need translation and waste tokens in a glossary.

It also swings between translation styles. Sometimes it preserves abbreviations like K2, and sometimes it expands them like V3.2. For example, for RNN tree structure in File 7, both V3.2 and K2 translated it simply as “RNN tree structure,” while V3 expanded it to “recurrent neural network tree structure,” which feels verbose. Still, it has places where it is better than V3.2. In File 4, V3.2 missed informal in cultural and informal institutions and produced “culture and institutions,” while V3 rendered the full meaning as “cultural and informal institutions.”

3. Do You Really Need Terminology Extraction?

#

Some people may ask whether turning off terminology extraction saves money. Yes, it does. But I do not recommend it, because the quality loss is large.

BabelDOC’s online feedback data shows that with automatic extraction enabled, one page averages around 7000 tokens; with it disabled, one page averages around 4500 tokens. In other words, spending only 2500 extra tokens per page can greatly improve translation accuracy.

As for token structure, when automatic term extraction is disabled, input-to-output tokens are roughly 3:1. When automatic term extraction is enabled, the ratio is roughly 4:1.

4. Summary and Recommendations

#

After this round of tests, the conclusion is clear.

Kimi K2 is currently the best model for terminology extraction. It can look stiff and preserve more English, but it rarely makes factual mistakes. It will not translate Company A as Company B or reverse academic concepts. For paper translation where accuracy matters, “not being wrong” is more important than “sounding nice.”

DeepSeek V3.2 produces the most fluent and natural Chinese, but it does over-interpret. However, its strong cache mechanism makes its cost-performance so attractive that it is hard to ignore.

💡 About DeepSeek’s “Disk Cache”

#

Before recommending concrete combinations, we need to mention DeepSeek’s official API context disk cache.

Simply put, DeepSeek automatically caches your input content on disk. This feature is enabled for all users automatically. As long as your input prefix matches a previous one and is longer than the very low threshold of 64 tokens, the system reads from cache instead of recalculating, at a very low price of only 0.2 RMB per million tokens.

Recommended Combinations

#

Note: based on the test data, the token input volume in the terminology extraction stage is about 2 times that of the translation stage, while output is only about 1/2 of the translation stage. Use your best model where it matters most.

Plan A: Quality First

#

• Terminology extraction: Pro/moonshotai/Kimi-K2-Instruct-0905 on SiliconFlow
• Body translation: deepseek-chat from DeepSeek official
• Assessment: K2 controls terminology accuracy rigorously, while DeepSeek ensures fluent body translation. This is currently the most stable pairing.

Plan B: Maximum Value

#

• Terminology extraction: still use Pro/moonshotai/Kimi-K2-Instruct-0905 on SiliconFlow
• Body translation: use Alibaba Cloud qwen-plus
• Assessment: Qwen-plus provides solid translation quality, and Alibaba Cloud’s low pricing keeps cost excellent.

Plan C: Lazy Setup, Very Low Price

#

• All stages: use DeepSeek official deepseek-chat throughout
• Assessment: V3.2 may occasionally over-paraphrase, but configuration is simplest because no platform switching is needed. Most importantly, the entire workflow benefits from DeepSeek’s context cache. Once the cache hits, cost drops to about one tenth and becomes almost negligible.

Rate-Limit Configuration Suggestions

#

To avoid errors from excessive concurrency, use the following rate-limit settings as a reference:

Debugging suggestion: keep Pool Worker fixed at 200 for terminology extraction and 500 for body translation. QPS is more affected by network fluctuation, so start testing from 100. If the console reports 429 errors (Too Many Requests), lower it slightly until the errors stop.

5. Online Experience

#

If you are an Immersive Translate member, you can now directly try Kimi + Qwen and Kimi + DeepSeek combination services. Click for details

Click here to quickly try BabelDOC.

Credits

#

Blog copy: Rosetears & awwaawwa Translation testing: Rosetears & awwaawwa Video recording: Rosetears Token-usage statistics: Immersive Translate Special thanks: awwaawwa & Immersive Translate

References

#

[1]MAZZOCCA C, ACAR A, ULUAGAC S, et al. A Survey on Decentralized Identifiers and Verifiable Credentials[J/OL]. IEEE Communications Surveys & Tutorials, 2025: 1-1. DOI:10.1109/COMST.2025.3543197. [2]REICHHELD F F. The One Number You Need to Grow[J]. Harvard business review, 2003, 81(12): 46-55. [3]DUL J. A different causal perspective with Necessary Condition Analysis[J/OL]. Journal of Business Research, 2024, 177: 114618. DOI:10.1016/j.jbusres.2024.114618. [4]DUL J, HAUFF S, BOUNCKEN R B. Necessary condition analysis (NCA): review of research topics and guidelines for good practice[J/OL]. Review of Managerial Science, 2023, 17(2): 683-714. DOI:10.1007/s11846-023-00628-x. [5]CHEN C. Science Mapping: A Systematic Review of the Literature[J/OL]. Journal of Data and Information Science, 2017, 2(2): 1-40. DOI:10.1515/jdis-2017-0006. [6]MUKHERJEE D, LIM W M, KUMAR S, et al. Guidelines for advancing theory and practice through bibliometric research[J/OL]. Journal of Business Research, 2022, 148: 101-115. DOI:10.1016/j.jbusres.2022.04.042. [7]SOCHER R. RECURSIVE DEEP LEARNING FOR NATURAL LANGUAGE PROCESSING AND COMPUTER VISION[J].