Four-Model Translation Showdown: Week 48 Quality Evaluation, passthrough Leads with 8 Points

This week, 431 translation tasks were completed by 4 models. Three samples underwent multi-model blind evaluation, with passthrough the best overall at an average score of 8/10.

This week, 431 translation tasks were completed by 4 models. 3 samples were selected for multi-model blind comparison; best overall: passthrough (average score 8/10).

This Week's Translation Statistics

ModelLanguageTranslation VolumeAverage TimeAverage Quality Score
deepseek-v4-flashen7720.9sNot rated
claude-sonnet-4.6ja21547.4sNot rated
passthroughen1330sNot rated
native-englishen1-Not rated
deepseek-v4-flashzh15.9sNot rated
deepseek-v4-flash:backtranslateen47.5sNot rated

Sampled Comparative Evaluation

Evaluation 1: Apple's In-House AI Agent Revealed: Let It Operate Your Mac for You

ModelAccuracyFluencyTerminologyReadabilityTotal Score
passthrough98978
deepseek-v4-pro68787
gpt-o379898

passthrough

✓ Accurately preserves the core information of the original text; for example, “It browses the web, works across apps” directly corresponds to the description of the agent’s functions, with no obvious additions.

✗ The end of the text is severely truncated: after “and it chooses the right tool for each step. It does all this privately and with the polish people expect from Apple.” it cuts off directly, affecting completeness.

deepseek-v4-pro

✓ Paragraph transitions are smooth; for example, the transition from “answering questions” to “completing tasks” is natural, with clear logic.

✗ Over-adds content not present in the original, such as the specific reporter name Lucas Ropek, the date November 20, 2026, and expansions such as “digital colleague”; this is excessive free translation.

gpt-o3

✓ The language is natural and idiomatic; “digital colleague” and “adjusting its strategy when pro” are fluent and highly readable.

✗ It also adds the reporter’s name and date not mentioned in the original, which is unnecessary expansion; compared with A, it is more complete but deviates from the original.

Conclusion: Version A is the most faithful to the original but incomplete due to truncation; B and C are more fluent but both over-add content. It is recommended to prioritize the complete version of A or lightly polish based on A.

Evaluation 2: Nadella's Prediction: AI Agents Will Reshape the Internet Economic Landscape

ModelAccuracyFluencyTerminologyReadabilityTotal Score
passthrough98978
deepseek-v4-pro68797
gpt-o369897

passthrough

✓ “The current aggregators who have a business model which was for the human web will get, in some sense, reshuffled” directly preserves Nadella’s original words, with high accuracy.

✗ The sentence at the end of the article is suddenly truncated: “another shopping platform such as Shopify, could allow them, and” is incomplete.

deepseek-v4-pro

✓ Adds an “Editor's Note” paragraph, summarizing the business logic as “attract eyeballs, then sell them to advertisers,” enhancing readability.

✗ Adds a large amount of explanatory content not in the original, such as metaphors like “digital agents” and “earthquake,” which is excessive free translation.

gpt-o3

✓ “use free content to attract attention, then sell that attention to advertisers” is natural and fluent, with clear logical connections.

✗ It also adds an “Editor's note” and expanded paragraphs, deviating considerably from the original content.

Conclusion: Version A is most faithful to the original; the other two versions both clearly add content, so A is recommended first.

Evaluation 3: Cursor Key Leak Leads to Malicious Claude Code Extension Listing, 5,000 Developers Affected

ModelAccuracyFluencyTerminologyReadabilityTotal Score
deepseek-v4-flash87877
deepseek-v4-pro98988
gpt-o399999

deepseek-v4-flash

✓ The event timeline is described clearly; “On October 29, the malicious extension was listed on Cursor's VS Code extension marketplace cache” accurately restores the details of the extension’s listing.

✗ The paragraph structure is somewhat loose; the text goes directly into the body under the title without a transition, and the final sentence is truncated, making it incomplete.

deepseek-v4-pro

✓ Terminology consistency is good; “historical AI coding-agent trajectories” accurately corresponds to “AI coding-agent traces,” and the title translation is complete and standard.

✗ Some sentences are too long and require pauses when reading; “The attacker used the key to copy historical AI coding-agent trajectories from a read-only storage bucket” is somewhat cumbersome.

gpt-o3

✓ The language is the most natural and fluent; the heading “Fact Reconstruction” is concise and professional, and “the attackers had already obtained extension marketplace distribution credentials” is logically smooth.

✗ There are no obvious flaws compared with version B, but “user traces” is slightly more ordinary than B’s “trajectories,” and the terminology was not further optimized.

Conclusion: Version C (gpt-o3) is the best overall, leading in fluency and readability; version B is next, with more rigorous terminology handling; version A is slightly weaker in structure and completeness.