This week, 402 translation tasks were completed by 4 models. 3 articles were sampled for a multi-model blind evaluation comparison; the overall best was gpt-o3 (average score 8.7/10).
This Week's Translation Statistics
| Model | Language | Translation Volume | Avg. Time | Avg. Quality Score |
|---|---|---|---|---|
| deepseek-v4-flash | en | 79 | 23.2s | Not rated |
| claude-sonnet-4.6 | ja | 200 | 47.6s | Not rated |
| deepseek-v4-flash:backtranslate | en | 3 | 37.7s | Not rated |
| passthrough | en | 116 | 0s | Not rated |
| native-english | en | 2 | - | Not rated |
| deepseek-v4-flash | zh | 2 | 5.4s | Not rated |
Sampled Comparative Evaluation
Evaluation 1: Google DeepMind and Singapore Government Join Forces: How On-Device AI Identifies Scams Without Touching User Data
| Model | Accuracy | Fluency | Terminology | Readability | Total |
|---|---|---|---|---|---|
| claude-sonnet-4.6 | 8 | 7 | 9 | 8 | 8 |
| deepseek-v4-pro | 9 | 8 | 9 | 8 | 8 |
| gpt-o3 | 9 | 9 | 8 | 9 | 9 |
claude-sonnet-4.6
✓ Terminology consistency is good; AI terms such as 「エンドツーエンド暗号化」 and 「オンデバイス推論」 are accurate and consistent.
✗ Some sentences are slightly stiff; 「分析処理はユーザーデバイスの内部で完結し」 reads like a translation.
deepseek-v4-pro
✓ Fluency is relatively good; 「2段階の直列パイプライン」 is expressed naturally and professionally.
✗ The title phrase 「オンデバイスAIはいかにして」 has a slightly bookish translation tone, making readability a bit weaker.
gpt-o3
✓ Overall readability is best; 「端末側推論は、技術的にこの矛盾を断ち切る」 connects smoothly and naturally.
✗ Some terminology differs slightly from A and B; mixing 「端末側」 and 「オンデバイス」 reduces consistency somewhat.
Conclusion: The three versions are close in overall quality. gpt-o3 is slightly better in fluency and readability; the Claude version is the least complete due to truncation, so the gpt-o3 version is recommended.
Evaluation 2: American Express Introduces New Framework: Preventing Unauthorized Purchases by AI Agents
| Model | Accuracy | Fluency | Terminology | Readability | Total |
|---|---|---|---|---|---|
| claude-sonnet-4.6 | 9 | 8 | 9 | 8 | 8 |
| deepseek-v4-pro | 9 | 9 | 9 | 9 | 9 |
| gpt-o3 | 8 | 8 | 8 | 8 | 8 |
claude-sonnet-4.6
✓ The explanation of the 「代理取引認証」 mechanism is relatively complete, citing a specific rule example such as 「1回の注文が200ドルを超える場合は自動決済を行わない」.
✗ The end of a paragraph is abruptly truncated, affecting readability; the final paragraph of the original is not fully presented.
deepseek-v4-pro
✓ Terminology is used consistently and naturally; expressions such as 「エージェント認可検証」 and 「認可範囲」 are clear, and the logical flow is smooth.
✗ Some sentences are a little long and read as slightly dense, but overall they remain natural.
gpt-o3
✓ The description of AI misjudgment scenarios is specific, such as 「ユーザーが単に価格を確認していただけなのに注文まで行ってしまう」.
✗ The content provided by this version is incomplete, and the ending cuts off abruptly, affecting the overall assessment.
Conclusion: Version B is the best overall, with a balance of accuracy, fluency, and readability; A and C are slightly behind due to incomplete content, so B is recommended.
Evaluation 3: DeepSeek V4 Pro Plunges 18.1 Points on Material Constraints, Yet Code Execution Scores a Perfect 100
| Model | Accuracy | Fluency | Terminology | Readability | Total |
|---|---|---|---|---|---|
| deepseek-v4-flash | 8 | 8 | 9 | 7 | 8 |
| deepseek-v4-pro | 9 | 8 | 9 | 8 | 8 |
| gpt-o3 | 9 | 9 | 9 | 9 | 9 |
deepseek-v4-flash
✓ Terminology consistency is good; 「Smoke evaluation」 retains the professional evaluation name directly without over-translation.
✗ The content is truncated; 「rather than one dimension c」 is clearly incomplete, affecting readability.
deepseek-v4-pro
✓ It accurately conveys the meaning of “plunges 18.1 points”; “Plunges 18.1 Points” is an apt wording.
✗ It also has truncation issues; 「systematic deg」 is unfinished, posing a risk of omission.
gpt-o3
✓ Fluency is best; the title translation 「Sampling Variance Rather Than Model Degradation」 is natural and professional.
✗ It is also truncated; 「there has been n」 is incomplete, but the overall structure is better than A and B.
Conclusion: Version C is the best overall, leading in fluency and readability; A and B are slightly behind due to truncation issues, but all three have relatively high accuracy and terminology consistency.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接