Release-note translation

DeepSeek V4 Pro vs Qwen 3.7 Max
for translating release notes

This page compares two current models on one job: translating release notes and update announcements for regional teams and customers. It covers translation fidelity, terminology control, cost and prompting, and ends with a fair way to test both.

Sep 8, 2026 · 12 min read

The bottom line
Qwen for fidelity and DeepSeek cost

Qwen 3.7 Max has the stronger published case for keeping technical meaning intact across languages. DeepSeek V4 Pro is the stronger case on price, current release status and open weights.

That split rests on Alibaba's own multilingual translation benchmarks8, DeepSeek's published token prices26, and an independent speed measurement910, not on a dedicated release-note translation test, since none exists publicly for these exact models.

In a staged workflow, many teams route customer-facing translation through Qwen 3.7 Max first, followed by bilingual review, and send high-volume internal drafts to DeepSeek V4 Pro, off-peak where the schedule allows. Either way, the final release still needs a human check of every version number, identifier and modal word against the source8.

Who this is for
Which regional teams this fits

Start with Qwen01

Localisation teams

You own the translation pipeline for release notes and app strings. Qwen 3.7 Max's published multilingual evidence is the closest thing to a translation-specific result, so it is the draft to start from before a bilingual review pass.

Add guardrails02

Technical writers

You write the English release notes that get translated downstream. Give either model an explicit list of what must survive unchanged: version numbers, API names, flags and modal words like may and will.

Watch verbosity03

Product marketers

You care that a translated announcement does not quietly promise more than the English original. Qwen's tendency to run verbose is worth checking here, since extra wording can read as an unsupported claim.

Run off-peak04

Regional operations teams

You push the same release across many markets on a schedule. DeepSeek V4 Pro's off-peak rate is the lowest option here, and the savings can fund the human check every region still needs.

What we compared
Translation accuracy not the app

This page compares the two models through their API in one neutral setup, not one model translating inside a CMS or localisation platform against the other inside a different tool.

The parts that matter for this job are keeping version numbers, feature names, API names, code spans and dates unchanged, preserving modal words like may and will, following a supplied glossary, and not turning a precise claim into something vaguer or friendlier. Official docs come first, then Alibaba's own translation benchmarks and an independent speed measurement.

We left tools out of the spec table on purpose. A translation-management platform, a glossary manager or a CMS plugin depends on the app around the model, so the same model can behave differently in a chat product, the API or a workspace. Judging those here would compare software, not translation.

Specs at a glance
The translation-relevant numbers

The model facts that actually affect translating a release note. Tool features are left out, since they change with the app around the model.

Spec
DeepSeek V4 Pro
Qwen 3.7 Max
Why it matters
Context window
1,000,000 tokens
1,000,000 tokens
Room to hold a product glossary, style guide and earlier releases alongside the current one46
Max output
384,000 tokens
131,072 tokens
DeepSeek can return a longer combined translation batch in one response46
List price (peak)
$1.32 in / $3.96 out per million
$1.65 in / $4.951 out per million (US/global)
DeepSeek's peak rate already undercuts Qwen's standard rate26
List price (off-peak)
$0.66 in / $1.98 out per million
No published off-peak discount
Scheduling a high-volume translation run off-peak roughly halves DeepSeek's cost2
Cached input price
$0.022 off-peak / $0.044 peak per million
$0.33 per million
Repeated glossary and style text across a release package costs much less to resend with DeepSeek26
Weights
Open weights
Proprietary
Only matters if self-hosting or direct control over the model is a requirement46

Figures from DeepSeek and Alibaba Cloud documentation, checked September 2026. Qwen's Singapore-scoped deployment lists $2.50 input and $7.50 output per million tokens instead of the US/global rate shown here6.

Head to head
Where the published evidence points

The clearest evidence favors Qwen on translation-specific measures, but it compares against an older DeepSeek preview, not the current release. This is the main analysis, row by row.

Job
Better choice
Why the edge exists
Best evidence
Broad multilingual translation accuracy
Qwen 3.7 Max, directional
WMT24++ covers 55 languages and dialects with human references and post-edits, the most task-relevant published score here. The DeepSeek figure is for its April preview, not the current release, so the edge is directional.
Qwen scored 85.8 against 82.2 for the DeepSeek preview on WMT24++8
Keeping detailed instructions across languages
Qwen 3.7 Max, slight edge
MaXIFE tests 1,667 verifiable instruction tasks across 23 languages. The gap is small, so the fair reading is that both are capable rather than one being clearly ahead.
Qwen scored 89.2 against 88.9 for the DeepSeek preview on MaXIFE8
Long release packages and translation memory
Qwen 3.7 Max, tentative
Both expose a 1M-token context, but Qwen's retrieval score at 128K is well ahead. Again, the DeepSeek side is the older preview, not the current release.
Qwen scored 90.4 against 74.4 for the DeepSeek preview on MRCR-v2 at 128K8
Terminology and format control
Tie
Both APIs return valid JSON, so each paragraph can be paired with a translation and a flagged-terms list. Neither guarantees correct terminology without a supplied glossary to check against.
Both vendors document JSON output support, with no exact-version terminology benchmark36
Cost per translation run
DeepSeek V4 Pro
DeepSeek's peak rate already sits below Qwen's standard rate, and its off-peak rate is far lower again.
DeepSeek lists $0.66-$1.32 in and $1.98-$3.96 out against Qwen's $1.65 in and $4.951 out per million tokens26
Current model lifecycle
DeepSeek V4 Pro
DeepSeek's alias resolves to the August 13, 2026 release. Qwen 3.7 Max remains callable but still points to its May 20 snapshot, and Alibaba's recommended list now leads with Qwen 3.8 Max instead.
DeepSeek's current alias is the August 13 build, Qwen 3.7 Max is no longer Alibaba's recommended Max model17
Response speed
Qwen 3.7 Max
Independent measurement found Qwen answering markedly faster, which matters when a team is iterating on many release notes in one sitting.
Artificial Analysis measured about 200 output tokens per second for Qwen against about 80 for DeepSeek V4 Pro910

Better-choice calls map to dimensions Alibaba and independent testers actually measured. Where the DeepSeek side of a comparison is its older preview rather than the current release, the row says so.

How to test
A fair test on your own release notes

A useful test stays boring. Same prompt, same glossary, same source text, same target language, same thinking setting. Then judge what actually matters: did every number, name and modal word survive, and how much bilingual correction it needed.

Sample01

Pick five real releases

Cover the range: a normal feature announcement, a deprecation or breaking-change notice, a security or permissions update, a release full of code identifiers and version strings, and a deliberately difficult note with negation, conditions and modal verbs.

Prompt02

Use one shared prompt

Run both APIs with the same system prompt, glossary, source text, target language, temperature and thinking setting. Do not edit either output before scoring it.

Setup03

Match the run setup

Test the API configuration that will actually run in production, since API and chat-product outputs can differ. Start with thinking disabled on both models for a first pass, and save a thinking-enabled second pass for auditing omissions.

Scoring04

Score blind and bilingual

Check numbers, dates, versions, product names, code spans, modal words and glossary terms against the source, and confirm no condition or platform was dropped. Use blind bilingual review for anything customer-facing, and record how much hand correction each output needed.

What the evidence shows
Strong for Qwen but not settled

No public benchmark tests release-note translation directly, so the best evidence is a mix. Here is what each source actually helps judge.

Source
What it measures
What it suggests
How to weigh it
WMT24++
English-to-target translation across 55 languages and dialects, with human references and post-edits
Qwen 3.7 Max scores 85.8 against 82.2 for the DeepSeek preview
The strongest translation-specific evidence here, but it does not isolate technical terminology or use the current DeepSeek release8
MaXIFE
Verifiable instruction following across 23 languages
Qwen leads by only 0.3 points, 89.2 to 88.9
Too small a gap to call a clear winner. Both look capable8
MMLU-ProX and NOVA-63
Multilingual reasoning and native-language knowledge, not translation
Qwen scores strongly in Alibaba's own table
Useful for general multilingual comprehension, not proof of translation fidelity8
DeepSeek's August model card
Agent and software-engineering results, including Terminal Bench and DeepSWE
Confirms the current release differs materially from the April preview
Shows real improvement exists, but it cannot substitute for a translation-specific score45
Artificial Analysis
Composite intelligence, speed and price across many task types
Qwen is fast but relatively verbose, DeepSeek shows a large jump from its own preview
A broad leaderboard, not a terminology-fidelity measurement910

No independent, exact-version translation shootout exists yet for these two models. The comparison that shows a Qwen lead used DeepSeek's older April preview, not the current DeepSeek V4 Pro release85.

How to prompt each one
State the invariants explicitly

The best prompt is not the same for both. Naming exactly what must survive the translation does more for accuracy than either model's default behavior.

DeepSeek's own prompt guide for translating its harness documentation asks for the same shape: preserve code spans, API names, paths, version numbers, dates and Markdown structure exactly, apply a supplied terminology table, and flag an ambiguous term for human review instead of guessing at it. That guide is written for translating DeepSeek's own repository docs. The same prompt shape carries over to release notes11.

Qwen 3.7 Max responds well to a similar shape. Match each source paragraph to one translated paragraph, and keep negation, conditions, numbers and identifiers intact. Alibaba's own Qwen-MT guidance documents a comparable way to control terminology: supply a glossary through term intervention so brand names and technical terms translate consistently, alongside domain prompting and translation-memory examples. That guidance is written for the dedicated Qwen-MT model. The same pattern carries over to Qwen 3.7 Max: ask it to flag anything it could not match against your glossary12.

For a first pass on either model, start with thinking mode off. Chain-of-thought prompting has been shown to push machine translation toward an overly literal reading, so save the deeper reasoning mode for a second-pass audit that checks for omissions and contradictions13.

A DeepSeek V4 Pro prompt: explicit invariants and a review field

Translate the release note into German.
Preserve all code spans, API names, paths, version
numbers, dates and Markdown structure exactly.
Use the glossary exactly.
Do not add, omit, summarize or soften any claim.

If a technical term is ambiguous, keep the source term
and list it under `review_required`.

Return JSON with `translation` and `review_required`.

A Qwen 3.7 Max prompt: paragraph alignment and unresolved terms

Translate the release note into Japanese for enterprise
customers. Match each source paragraph to one translated
paragraph.

Preserve technical scope, negation, conditions, numbers,
identifiers and Markdown. Apply the glossary exactly.
Do not explain or improve the source.

Return only the translation and a list of unresolved
terminology.

Weak spots
And how to fix them

Neither model is proven on this exact job yet. The useful question is where each one is likely to add review work, and what to change in the prompt or workflow.

Model
Weak spot
What it looks like
How to fix it
DeepSeek V4 Pro
No public translation score for the current release
The August model card reports agent and coding gains, not an updated translation benchmark, so its real improvement on this exact job is unmeasured.
Treat it as unproven until it passes your own test set, with sentence-level source alignment and a protected-token list45.
DeepSeek V4 Pro
Rewriting can shift scope on a bare translate instruction
Natural rewriting can add tone, drop a qualifier or soften a claim when the prompt does not forbid it.
Explicitly forbid summarizing, explaining or softening, and route ambiguous terms into a separate review field11.
Qwen 3.7 Max
Runs verbose
Independent testing describes it as relatively wordy, and extra wording in a release note can read as an unsupported explanation or promise.
Require translation-only output with paragraph alignment, and automatically check numbers and protected identifiers against the source9.
Qwen 3.7 Max
No longer Alibaba's lead Max model
The qwen3.7-max alias still resolves to its May 20 snapshot, while Alibaba's recommended list now leads with Qwen 3.8 Max.
Pin the dated snapshot for reproducibility, or add Qwen 3.8 Max to the same test67.
Both
Fluent output can hide a dropped detail
A glossary can conflict with natural phrasing, and smooth-reading text makes an omission easy to miss.
Keep a maintained per-language glossary, run automated token and number checks, and require bilingual human approval before anything customer-facing ships.

Which one to choose
Start from your priority

One question first. Is the priority fidelity for customers, volume and cost internally, or self-hosting? Then follow the branch that matches most of your releases.

What matters most for this release? Customer-facing release notes High-volume internal drafts Self-hosting or open weights required Breaking changes or security notices Qwen 3.7 Max DeepSeek V4 Pro DeepSeek V4 Pro Either plus human review Verify against source

A starting point, not a rule. Test on your own release notes before you commit

Recommendations
Pick by your release profile

If most of your release notes reach paying customers directly, start with Qwen 3.7 Max for the first draft. Its published multilingual evidence is the closest thing to a translation-specific result8, but always follow it with a bilingual review pass before anything ships.

If the bulk of the work is high-volume internal drafts, run DeepSeek V4 Pro off-peak where your schedule allows. Its rate is well below Qwen's even at peak hours, and the savings can fund the human review step that both models still need26.

If self-hosting or control over the model weights is a requirement, DeepSeek V4 Pro is the only option here with open weights, since Qwen 3.7 Max is proprietary6. For breaking changes, security advisories or anything with legal weight, use either model only as a draft and require a human reviewer to check every number, identifier and modal word against the source before release.

One honest limit: if you are a single localiser working alone on one language pair with no team to split review work across languages, Playgram buys you less. The comparison above still helps pick a model. The shared workspace Playgram offers below is built around a team splitting the review work.

Bottom line
Neither claim is a same-version test

Qwen 3.7 Max has the stronger published case for keeping meaning intact across languages. DeepSeek V4 Pro has the stronger case on price, current release status and open weights.

The limits matter here more than usual. The clearest evidence for Qwen came from a comparison against DeepSeek's older April preview, not the current August release, so the real gap between the two current models is unmeasured85. Most of the reported numbers are vendor-published, and no public benchmark isolates whether either model quietly turns a precise release-note claim into something vaguer. Benchmarks also average across languages, which can hide a weak result in one specific target language.

The safest final step is to test the shape of your own release notes, not a generic translation prompt from the internet. A fair test needs the same setup for both models: the same source text, the same prompt and the same place to run them, so the result reflects the models and not the tool around them. In practice that is harder than it sounds, since most teams end up running one model in one app and the other in a different one, on two separate subscriptions, which tilts the comparison before the first translation comes back. The cleaner the setup, the more the difference you see is really DeepSeek V4 Pro vs Qwen 3.7 Max, and not just which one happened to be easier to reach that day.

Test both in one workspace
Right here inside Playgram

That's the practical case for the setup just described, and it's also how the day-to-day work gets easier. When both models sit in one workspace, a localisation team can send the same release note to each, compare the two translations side by side, and hand a draft from one model to the other without setting up the glossary and context again.

Playgram lets you run that same comparison directly. Paste the release notes for translation once, put them in front of the latest DeepSeek and Qwen models, and keep the conversation going with either one without re-pasting the glossary or starting over for a second opinion.

The same memory carries across the team too, not just this one comparison, over the latest GPT, Claude, Gemini and Grok models and many more, all in one place14. The line-up is curated, so retired models are turned off and new ones are added as they ship.

Team memory

Shared across everyone and every model.

Project memory

Scoped to a campaign or document set.

Personal memory

Your own working style, kept private.

Fair pricing
Pay per usage, not per seat

Upgrade as needed, and only pay for what you actually use

Save ~17% with the annual plan

Pro

$50/ month

Perfect for small and medium teams

Unlimited users & infinite memory

Multi-LLM chats

Granular access control to models

EU data residency

Get started

Ultra

$200/ month

Best for large, growing teams

Unlimited users & infinite memory

Multi-LLM chats

Unlimited use of DeepSeek V4 Flash

Granular access control to models

Choose US or EU data residency

Get started

Enterprise

Get in touch

Unlimited Credits

For organizations with advanced needs

Unlimited users & SSO

Priority Support

Unlimited use of DeepSeek V4 Flash

Granular access control to models

Choose US or EU data residency

Book a call

30-days money back guarantee

Frequently asked
questions

Neither has a public benchmark built for exactly this job, but the closest evidence favors Qwen 3.7 Max. Its 85.8 score on WMT24++, a 55-language translation benchmark with human references, is the best available translation-specific signal. The comparison used DeepSeek's older April preview rather than the current DeepSeek V4 Pro release, so treat the lead as directional and check both against your own protected-term list.

Untested, as far as the public record shows. DeepSeek's August model card reports large gains on agent and coding benchmarks like Terminal Bench and DeepSWE, but no updated translation score. The 82.2 WMT24++ figure often cited against Qwen belongs to DeepSeek's April preview, not the current build, so there is no clean evidence either way for the model actually in production today.

DeepSeek V4 Pro's off-peak rate is $0.66 input and $1.98 output per million tokens, roughly half its peak rate of $1.32 and $3.96. Qwen 3.7 Max does not publish an equivalent off-peak discount, listing $1.65 input and $4.951 output per million tokens at its US and global rate. For a team translating the same release into many languages on a schedule, timing the run can be worth more than the model choice.

Both can, if the prompt asks for it. DeepSeek's own documentation-translation prompt guide has the model flag an unclear term for human review rather than guessing, and Qwen-MT's terminology feature lets you supply a fixed glossary so a term either matches it or needs a human look. Neither model does this automatically on a bare translate this instruction.

Yes, with either model, if the prompt does not forbid it. Independent testing describes Qwen 3.7 Max as relatively verbose, and extra wording in a release note can slide into an unsupported explanation or promise. DeepSeek's natural rewriting can drop a qualifier or soften a claim on a bare translate instruction. Explicitly ban summarizing, explaining and softening, and check that modal words like may and will survive unchanged.

No. Run the same prompt on both and compare the answers, or switch between them mid-conversation. You choose after reading both answers instead of guessing up front.

Related comparisons

GPT-5.6 Sol vs Qwen 3.7 Max for multilingual writingClaude Sonnet 5 vs DeepSeek V4 Pro for interview synthesisClaude Sonnet 5 vs GPT-5.6 Sol for policy translationGrok 4.5 vs Kimi K3 for marketing copy translation

One release note two models
One glossary one memory

Send the same release note and glossary to the latest DeepSeek and Qwen models, keep the terminology in one place, and see which translation needs fewer fixes. Set it up in a minute.

Get startedSee the pricing