{"slug": "translation-specialists-vs-frontier-llms-a-reference-table", "title": "Translation Specialists vs Frontier LLMs: A Reference Table", "summary": "Tencent's Hy-MT2 family of open-weight translation specialists, hosted on OpenRouter since 19-20 August 2026, outperforms frontier LLMs on WMT25 GEMBA and XCOMET-XXL benchmarks but trails on FLORES-200, according to Tencent's own evaluations in arXiv 2605.22064. Hy-MT2-30B-A3B scored 84.34 on WMT25 GEMBA versus GPT-5.5's 83.29 and Gemini 3.1 Pro's 82.23, while on FLORES-200, Gemini 3.1 Pro scored 94.42 and GPT-5.5 94.16, ahead of Hy-MT2-30B-A3B's 93.85. The models were open-sourced on 21 May 2026, and OpenRouter lists them with 8,192-token context despite the model's 262,144-token capability.", "body_md": "Specialist translation models and frontier LLMs are rarely compared on the same page, because the dimensions a localisation buyer decides on — language coverage, context window, licence, price per unit and evaluation evidence — live across a dozen vendor pages and one dense academic table. This reference collects them into 17 rows, as of 22 August 2026, with the benchmark name attached to every score and the unit attached to every price.\n\nThe occasion is Tencent’s Hy-MT2 family, which OpenRouter added hosting for on 19–20 August 2026, three months after the models were open-sourced on 21 May 2026. Its paper publishes a comparison table in which a 30B mixture-of-experts translation specialist scores above GPT-5.5 and Gemini 3.1 Pro on one metric and below them on another. Both results are in the table below, with their metric names, because either one alone would mislead.\n\nWhat follows is the methodology, the category and metric definitions, the complete dataset, one chart rendered from it, and the cells we deliberately left empty. The prose is packaging; the table is the asset.\n\n- 0117 rows, three categories, one claim per cell.Seven open-weight MT specialists (Hy-MT2 ×3, NLLB-200-3.3B, SeamlessM4T v2 Large, TowerInstruct-Mistral-7B-v0.2, MADLAD400-3B-MT), six commercial API tiers (Google Cloud Translation ×3, Azure Translator S1, DeepL Developer and Growth) and four frontier LLMs as Tencent-run baselines.\n- 02The specialist-vs-frontier result runs both ways.In arXiv 2605.22064 Table 2, Hy-MT2-30B-A3B posts the best WMT25 GEMBA score of every compared system (84.34 vs GPT-5.5 T 83.29, Gemini 3.1 Pro T 82.23), and the specialists lead the WMT25 XCOMET-XXL column too (Hy-MT2-7B 63.86, 30B-A3B 62.89 vs Gemini 3.1 Pro T 57.58, GPT-5.5 T 56.68). The order flips on the same table’s FLORES-200 EN⇔XX column, where the XCOMET-XXL scores are 94.42 for Gemini 3.1 Pro T and 94.16 for GPT-5.5 T against 93.85 for the 30B-A3B, 93.52 for the 7B and 90.00 for the 1.8B.\n- 03Every benchmark figure here is Tencent’s own run.The frontier-LLM scores are Tencent’s reproduction of those systems on WMT25 and FLORES-200, not the vendors’ self-reported numbers and not an independent leaderboard. Read each score with that provenance attached.\n- 04Hy-MT2 launched 21 May 2026; OpenRouter listed it in August.OpenRouter’s created timestamps are 2026-08-19 14:13 UTC for the 7B, 2026-08-20 13:12 UTC for the 30B-A3B and 13:13 UTC for the 1.8B — a 90–91 day listing lag. OpenRouter serves 8,192 tokens of context; the model’s own config.json allows 262,144.\n- 05Prices keep their units; some cells stay empty.OpenRouter bills per million tokens, the commercial APIs per million characters, and the two are not converted here. OpenAI’s and Google’s official language counts for GPT-5.5 and Gemini 3.1 Pro, and DeepL’s Developer-tier overage rate, were not located on a vendor page and are marked as such.\n\n## 01 — MethodologyWhat was read, what was *inferred*, and from where.\n\nEach row rests on a primary file we opened: a Hugging Face model card, a `config.json`\n\n, a `LICENSE.txt`\n\n, the arXiv PDF’s full text, OpenRouter’s public models endpoint, or a vendor pricing page. Where a value was not read from such a file — a licence inferred from a sibling repo, a language count carried from the paper rather than the card — the block below says so. The aim is that a stranger can redo the work and land on the same numbers.\n\n**What was collected.** A 17-row comparison of machine-translation systems in three categories: (1) open-weight translation specialists, (2) commercial translation API tiers, (3) frontier general-purpose LLMs used for translation via prompting. Ten columns per row: system, maker, type, languages (count plus its source), context, licence and open-weights status, price with its own unit, evaluation evidence (benchmark, metric, score, who ran it), intended use, and source with date. **Data as of 22 August 2026.**\n\n**Sources, and what was read versus inferred.** *Read:* the full text of arXiv 2605.22064 (Tables 2, 3 and 5 plus the §3 prose that names each metric); the raw `LICENSE.txt`\n\non the Hy-MT2-30B-A3B and Hy-MT2-1.8B Hugging Face repos; the 30B-A3B repo’s `config.json`\n\nand README YAML; the model cards for NLLB-200-3.3B, SeamlessM4T v2 Large, TowerInstruct-Mistral-7B-v0.2 and MADLAD400-3B-MT; OpenRouter’s public `/api/v1/models`\n\nendpoint for all three Hy-MT2 listings; Google Cloud Translation and Azure AI Translator pricing pages via a rendered fetch (the Azure page’s static HTML serves placeholder prices; the render used carried a 2026-08-20 cache date); DeepL’s API plans page via JSON extraction of a snapshot cached 2026-08-21; Anthropic’s multilingual-support documentation; and the Gemini API models documentation — fifteen primary files and pages in all. *Inferred, not read:* the Hy-MT2-7B licence (taken from the family, its own LICENSE.txt was not opened); the language count on the 1.8B and 7B rows (the paper’s abstract states 33 languages for all three sizes; only the 30B-A3B README YAML was counted); the 262,144 context figure (read only on the 30B-A3B config). The active-parameter arithmetic behind the “A3B” name was not recomputed.\n\n**When.** All primary fetches were made in a single research session at the time of writing, and the dataset is stated as of 22 August 2026. Every fetched page is either dated on or before that day or is a stable reference page (pricing, model card, licence file) with no indication its content changed.\n\n**Excluded, and why.** ALMA (an earlier LLM-based MT fine-tune) — its maintenance status could not be confirmed within the research budget, so it was omitted rather than guessed. Amazon Translate and Yandex Translate — outside the brief’s research targets; omitted for scope, not because they do not exist. WMT24 and earlier test sets — only WMT25 is cited, because it is the benchmark the Hy-MT2 paper uses and its findings paper frames itself as correcting easy-test-set inflation. The paper’s other baselines (Qwen3.5, GLM5.1, Gemma 4, Tower-Plus-72B, TranslateGemma, Doubao, iFLYTEK) and the non-thinking-mode runs of the four frontier LLMs are not carried as rows; the non-thinking scores are quoted in prose where they change the reading.\n\n**Known limitations.** (1) Every Hy-MT2-versus-LLM score is Tencent’s own evaluation run, including the baselines; no independent reproduction was found. (2) DeepL’s overage rate for its free Developer tier could not be read from DeepL’s own rendered page; a third-party figure circulates and is flagged, not asserted. (3) Neither OpenAI nor Google publishes a fixed supported-language count for GPT-5.5 or Gemini 3.1 Pro that we could locate; those cells are left as not located. (4) An earlier automated summary of the same paper, produced during research, mislabelled which column was WMT25 and which was FLORES-200, so the benchmark-labelled numbers here were taken from the raw PDF text and cross-checked against the paper’s prose, which names each metric. (5) Benchmark scores for NLLB-200, SeamlessM4T, TowerInstruct and MADLAD were not re-extracted from their papers in this pass; those cells name the evaluation sets only. (6) Table 2 reports FLORES-200 in three direction settings — ZH⇔XX, EN⇔XX and XX⇔XX — alongside a Mandarin⇔Minority column. This table carries the EN⇔XX and XX⇔XX settings and omits ZH⇔XX and Mandarin⇔Minority; a figure quoted from one setting is not comparable with another.\n\n**To redo this.** Pull `https://openrouter.ai/api/v1/models`\n\nand filter `id`\n\nfor `hy-mt2`\n\n; fetch `https://arxiv.org/pdf/2605.22064`\n\nand read Table 2 (general translation) and Table 3 (domain and in-the-wild); open each Hugging Face repo’s raw README and LICENSE; open the Google Cloud, Azure and DeepL pricing pages with a JavaScript-rendering fetch. Roughly 90 minutes for one person.\n\n## 02 — Categories and MetricsThree kinds of system, three kinds of *score*.\n\nThe table mixes systems that are not normally compared, so the categories need stating. An **open-weight translation specialist** is a model trained or fine-tuned for translation whose weights can be downloaded and self-hosted; it typically accepts a sentence or paragraph and returns its translation, which is why context windows in this category are small. A **commercial translation API** is a hosted service billed per character, with no weights and no published benchmark score of its own. A **frontier LLM used for translation** is a general-purpose model prompted to translate; its language support is described qualitatively by its vendor, and its price is set for general use rather than for translation.\n\nThe Hy-MT2 paper scores general translation with three automatic metrics, and the distinction between them is the whole story of this dataset. In the paper’s own words, XCOMET-XXL is a reference-based metric (it compares the output to a human reference translation), CometKiwi is reference-free (it judges source and output without a reference), and GEMBA is an LLM-based metric (a large language model rates the translation). A system can lead on one and trail on another, and Hy-MT2-30B-A3B does exactly that.\n\n##### XCOMET-XXL\n\nScores how closely a translation matches a professionally produced reference. Rewards adequacy and fidelity. Hy-MT2-7B and 30B-A3B lead the WMT25 column; Gemini 3.1 Pro T and GPT-5.5 T lead the FLORES-200 EN⇔XX column, which scores about thirty points higher.\n\n##### CometKiwi\n\nEstimates quality from the source and the translation alone. Useful where no reference exists. Hy-MT2-7B (71.21) and 30B-A3B (71.08) lead the WMT25 column in Tencent's run.\n\n##### GEMBA\n\nA large language model scores overall quality and readability. The metric on which Hy-MT2-30B-A3B posts the best WMT25 score of every compared system in Tencent's run.\n\nTwo benchmark suites appear in the evaluation column. FLORES-200 is a multilingual sentence-level test set covering 1,056 translation directions across 33 languages, and Table 2 splits it by direction setting: EN⇔XX is English paired with each other language, while XX⇔XX is the paper’s average over all evaluated directions, EN⇔XX and ZH⇔XX included. The two settings score differently, so each cell here names the one it came from. WMT25 is the general machine-translation shared task of the Tenth Conference on Machine Translation, whose findings paper (Kocmi et al., 2025, published November 2025 under DOI [10.18653/v1/2025.wmt-1.22](https://doi.org/10.18653/v1/2025.wmt-1.22)) is titled “Time to stop evaluating on easy test sets,” and it covers 12 translation directions. Scores on WMT25 run far lower than on either FLORES-200 setting for every system where this table carries both; the columns are not comparable with each other.\n\n## 03 — The DatasetAll *seventeen* rows, one claim per cell.\n\nReading guide. **Languages** carries the count and where it came from. **Context** distinguishes what a host serves from what the architecture allows. **Price** states the vendor’s own unit — per million tokens on OpenRouter, per million characters on the commercial APIs — and is never converted. **Evaluation** gives benchmark, metric order (XCOMET-XXL / CometKiwi / GEMBA unless stated), score and who ran it. Frontier-LLM rows omit languages, context and price by design: those are general-purpose figures already tracked elsewhere, and restating them here would imply a translation-specific comparison that does not exist.\n\n| # | System (maker) | Type | Languages (source) | Context | Licence · weights | Price (unit stated) | Evaluation evidence (who ran it) | Intended use | Source · date |\n|---|---|---|---|---|---|---|---|---|---|\n| Open-weight translation specialists (rows 1–7) | |||||||||\n| 1 |\n|\n\n[Hy-MT2-7B](https://huggingface.co/tencent/Hy-MT2-7B)(Tencent)\n\n[Hy-MT2-30B-A3B](https://huggingface.co/tencent/Hy-MT2-30B-A3B)(Tencent)\n\n[config.json](https://huggingface.co/tencent/Hy-MT2-30B-A3B/raw/main/config.json)max_position_embeddings 262,144[LICENSE.txt](https://huggingface.co/tencent/Hy-MT2-30B-A3B/resolve/main/LICENSE.txt)read) · open weights incl. FP8[NLLB-200-3.3B](https://huggingface.co/facebook/nllb-200-3.3B)(Meta)\n\n[SeamlessM4T v2 Large](https://huggingface.co/facebook/seamless-m4t-v2-large)(Meta)\n\n[TowerInstruct-Mistral-7B-v0.2](https://huggingface.co/Unbabel/TowerInstruct-Mistral-7B-v0.2)(Unbabel)\n\n[MADLAD400-3B-MT](https://huggingface.co/google/madlad400-3b-mt)(Google)\n\n[cloud.google.com/translate/pricing](https://cloud.google.com/translate/pricing), rendered fetch, at time of writing[cloud.google.com/translate/pricing](https://cloud.google.com/translate/pricing), at time of writing[cloud.google.com/translate/pricing](https://cloud.google.com/translate/pricing), at time of writing[azure.microsoft.com/en-us/pricing/details/translator/](https://azure.microsoft.com/en-us/pricing/details/translator/), rendered fetch with a 2026-08-20 cache date[deepl.com/en/pro-api](https://www.deepl.com/en/pro-api), API tab; snapshot cached 2026-08-21[deepl.com/en/pro-api](https://www.deepl.com/en/pro-api), API tab; snapshot cached 2026-08-21[arXiv 2605.22064](https://arxiv.org/pdf/2605.22064)Table 2, v2 dated 2026-05-25Provenance note, stated here as well as in the methodology: **every score in rows 1–3 and 14–17, and the Microsoft-Translator figures quoted in row 11, comes from one evaluation run by Tencent** — the authors of Hy-MT2 — published as Table 2 of arXiv 2605.22064. The frontier-LLM numbers are Tencent’s reproduction of those systems, not OpenAI’s, Google’s, DeepSeek’s or Moonshot’s own reports, and no independent third party has re-run the comparison that we could find. Rows 16 and 17 list “proprietary in this comparison” because the open-weight status of DeepSeek-V4-Pro and Kimi K2.6 was not part of this research pass and is not asserted either way here.\n\n## 04 — One ChartTwo metrics, two *different* winners.\n\nThe chart plots the seven systems in the table that carry a WMT25 score — the three Hy-MT2 sizes and the four frontier baselines — on both WMT25 metrics, on a full 0–100 axis so that no gap is exaggerated. The values are the same ones printed in the table. The visual point is that one benchmark’s two metrics do not agree on a leader: GEMBA puts Hy-MT2-30B-A3B first and XCOMET-XXL puts Hy-MT2-7B first. On the reference-based metric both larger specialists sit above all four frontier baselines; on GEMBA the order is tighter, with GPT-5.5 T at 83.29 falling between the 30B-A3B’s 84.34 and the 7B’s 82.24.\n\n## 05 — Where Each Side WinsThe specialist leads WMT25; the frontier leads *FLORES-200* EN⇔XX.\n\n**Where Hy-MT2 wins.** On WMT25, Tencent’s Table 2 has Hy-MT2-30B-A3B at 62.89 / 71.08 / 84.34 and Hy-MT2-7B at 63.86 / 71.21 / 82.24 (XCOMET-XXL / CometKiwi / GEMBA). The 30B-A3B’s 84.34 GEMBA is the highest in the column: GPT-5.5 scores 83.41 in non-thinking mode and 83.29 in thinking mode, Gemini 3.1 Pro T 82.23, DeepSeek-V4-Pro T 81.99 and Kimi K2.6 T 81.68. The same benchmark’s reference-based column puts both larger specialists above all four frontier baselines as well: 63.86 for the 7B and 62.89 for the 30B-A3B against 57.58 for Gemini 3.1 Pro T, 56.68 for GPT-5.5 T, 54.83 for DeepSeek-V4-Pro T and 54.48 for Kimi K2.6 T. The paper states the GEMBA result in its own words:\n\n\"Hy-MT2-30B-A3B further achieves the best GEMBA score among all compared systems, surpassing Gemini 3.1 Pro T and GPT-5.5, indicating stronger overall translation quality and readability in challenging WMT settings.\"— Hy-MT2 technical report, arXiv:2605.22064 §3.2 (Tencent Hunyuan)\n\n**Where the frontier LLMs win.** The order reverses in the same table’s FLORES-200 EN⇔XX column, the higher-scoring of the two FLORES-200 settings carried here. Its reference-based metric reads 94.42 for Gemini 3.1 Pro T and 94.16 for GPT-5.5 T against 93.85 for Hy-MT2-30B-A3B, 93.52 for the 7B and 90.00 for the 1.8B — so the two largest frontier models beat every Hy-MT2 size on reference-based adequacy there, though the margin over the 30B-A3B is 0.57 of a point. These figures sit roughly thirty points above the WMT25 XCOMET-XXL column reproduced above because WMT25 and FLORES-200 are not test sets of comparable difficulty — the point the WMT25 findings paper’s title makes. On the XX⇔XX setting the paper puts the same point as a ratio: the three Hy-MT2 sizes reach 79.77, 86.89 and 87.47, which it reports as 89.9%, 97.9% and 98.6% of Gemini 3.1 Pro T. The specialist’s win is benchmark- and metric-specific, not general.\n\n**Mixed on the domain benchmarks.** Table 3 does not resolve the question either way, and it splits by metric. On DomainMTBench (finance, law, medical, technology, politics, education) Hy-MT2-30B-A3B averages 95.04 XCOMET / 93.73 GEMBA against Gemini 3.1 Pro T’s 94.50 / 94.64 and GPT-5.5 T’s 94.23 / 94.51 — the paper claims “the best XCOMET result among all compared systems” there, and the frontier models keep GEMBA. On WildMTBench, the in-the-wild set, the 30B-A3B reaches 89.87 / 89.25 and the paper claims “the best GEMBA score among all compared systems,” ahead of Gemini 3.1 Pro T’s 88.96. The pattern across the tables is that neither side sweeps: the specialist is strongest on the hardest general test set (WMT25) and on in-the-wild readability, while the largest frontier models hold FLORES-200 EN⇔XX and the domain-average GEMBA.\n\n##### Hy-MT2-30B-A3B leads\n\nHighest GEMBA in Tencent's Table 2. Next: GPT-5.5 83.41 (non-thinking), GPT-5.5 T 83.29, Gemini 3.1 Pro T 82.23.\n\n##### Hy-MT2-7B leads the rows carried here\n\nThe reference-based metric on the same benchmark: Hy-MT2-30B-A3B 62.89, Gemini 3.1 Pro T 57.58, GPT-5.5 T 56.68, DeepSeek-V4-Pro T 54.83, Kimi K2.6 T 54.48.\n\n##### Gemini 3.1 Pro T beats every Hy-MT2 size\n\nThe same table's higher-scoring FLORES-200 setting reverses the order: GPT-5.5 T 94.16, Hy-MT2-30B-A3B 93.85, Hy-MT2-7B 93.52, Hy-MT2-1.8B 90.00.\n\nFor anyone weighing these numbers, the caution from our [benchmark methodology guide](/blog/llm-benchmark-methodology-2026-contamination-leaderboard-guide) applies in full: a vendor-run comparison is evidence about the vendor’s setup, prompts and metric choices as much as about the models. That the same table contains both the flattering and the unflattering column is a point in the paper’s favour; that we found no independent reproduction of either column is the limit of what it can tell you.\n\n## 06 — Hy-MT2, Re-readThree sizes, one licence, two *context* numbers.\n\nBecause Hy-MT2 is the row most likely to be cited from this table, its cells were taken from the primary files rather than from coverage. The arXiv submission history dates v1 to 21 May 2026 07:00 UTC and v2 to 25 May 2026; the family therefore launched in May, and what happened in August is that [OpenRouter](https://openrouter.ai/api/v1/models) added hosting. Its `created`\n\ntimestamps read 1787148797 for `tencent/hy-mt2-7b`\n\n(2026-08-19 14:13:17 UTC), 1787231561 for `tencent/hy-mt2-30b-a3b`\n\n(2026-08-20 13:12:41 UTC) and 1787231581 for `tencent/hy-mt2-1.8b`\n\n(2026-08-20 13:13:01 UTC). An earlier sweep for this series enumerated only the 1.8B and 30B-A3B listings; the 7B row was present and had been created a day earlier. How to read an OpenRouter listing date in general is covered in our [model catalog literacy reference](/blog/ai-model-catalog-literacy-prices-dates-surfaces-2026); here the gap is 90 to 91 days.\n\n##### Hy-MT2-1.8B\n\nDense 1.8B. WMT25 50.30 / 64.59 / 70.36, against the Microsoft-Translator row Tencent ran in the same table at 42.01 / 60.21 / 67.63. The paper's abstract says it surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall — Tencent's own evaluation, not an independent one.\n\n##### Hy-MT2-7B\n\nDense 7B, priced identically to the 30B-A3B on OpenRouter — plausibly one backing provider, not asserted. Best WMT25 XCOMET-XXL (63.86) and CometKiwi (71.21) of the systems carried in this table.\n\n##### Hy-MT2-*30B-A3B*\n\nMixture-of-experts. config.json: 128 experts, 8 routed per token, MoE intermediate size 768 — consistent with the roughly-3B-active naming, not recomputed here. Best WMT25 GEMBA of all compared systems, per the paper.\n\n**Licence.** The raw `LICENSE.txt`\n\non the 30B-A3B repo reads “Hy-MT2-30B-A3B is licensed under the Apache License, Version 2.0,” and the [1.8B repo](https://huggingface.co/tencent/Hy-MT2-1.8B/resolve/main/LICENSE.txt) carries the equivalent file. The 7B’s licence file was not opened and is inferred from the family. Weights are on Hugging Face and ModelScope in BF16 and FP8. The GitHub repository at [Tencent-Hunyuan/Hy-MT2](https://github.com/Tencent-Hunyuan/Hy-MT2) also carries a separate `LICENSE-IFMTBench.txt`\n\n, almost certainly for a bundled benchmark dataset rather than the weights; it was not opened in this pass.\n\n**Context.** Two numbers are true at once. OpenRouter serves all three sizes with `context_length: 8192`\n\nand a 4,096-token completion cap — the size profile of a sentence- and paragraph-level translation model, and the reason these are not chat models. The 30B-A3B’s `config.json`\n\nsets `max_position_embeddings`\n\nto 262,144. The first is a serving decision; the second is the architecture’s ceiling. A team self-hosting the weights is bound by the second, not the first, though the paper publishes no long-document evaluation to say how quality holds up there.\n\n**Languages.** The paper’s abstract says all three sizes “support translation among 33 languages,” and Tencent’s materials describe the family as 33 languages plus dialect and minority-language pairs. The 30B-A3B [README](https://huggingface.co/tencent/Hy-MT2-30B-A3B/raw/main/README.md)’s YAML front matter lists 36 ISO-639 codes, the last four being bo (Tibetan), kk (Kazakh), mn (Mongolian) and ug (Uyghur); Traditional Chinese and Cantonese, also named by Tencent, have no separate code in that list and appear to be handled under zh. The table carries the paper’s 33 with the YAML count alongside it.\n\n**Quantisation.** The paper states twice that with AngelSlim’s 1.25-bit extreme quantisation the 1.8B “requires only 440 MB of storage and achieves a 1.5× inference speedup” — in the abstract, and again in the introduction, which names the baseline: 1.5× over the 4-bit quantised Hy-MT1.5 on an Apple A15, where the 4-bit build still needs more than 1 GB. What the paper does not publish is a quality score for that build — its quantisation ablation (Table 5) covers only BF16, FP8, Q4_K_M and 2-bit variants. Read 440 MB and 1.5× as storage and speed claims, not as quality results.\n\n## 07 — Price UnitsPer token and per character are *not* converted here.\n\nOpenRouter bills the Hy-MT2 listings per million tokens; Google, Microsoft and DeepL bill per million characters. A conversion between the two depends on the tokenizer and on the language — characters per token differ by several times between English and Chinese, Thai or Burmese — so any single exchange rate would be a fabrication dressed as arithmetic. The table keeps each vendor’s unit, and the chart below compares the five commercial tiers that share one — on a single basis, a million characters in and a million characters out where the tier bills both directions.\n\n#### Commercial translation API list prices · one million characters translated\n\nSource: vendor pricing pages, data as of 22 August 2026. Per million characters, output assumed equal in length to input so that the two-sided tiers are on the same basis as the one-sided ones. Two tiers are omitted: DeepL's Developer overage (not located on DeepL's page) and Google's Custom AutoML tier (volume-tiered, $80 down to $30 per M).Two things the bars do not say. Google’s Advanced LLM tier charges both directions, so a translation whose output is as long as its input costs the same $20 as the NMT tier — which is why the two sit level on the chart. On that same basis the Adaptive tier at $25 plus $25 is the most expensive of the five plotted, though Google’s Custom AutoML tier, left off because it is volume-tiered, starts higher still at $80 per million characters. And DeepL’s Growth overage applies only after the annual allowance is exhausted, so the headline $27.50 is a marginal rate on volume beyond roughly one million characters a month, not an entry price. A third-party claim that DeepL’s legacy “API Pro” plan ($5.49 a month plus $25 per million characters) was closed to new sign-ups in July 2026 circulates on pricing trackers; it is reported here, not verified against a DeepL page.\n\n## 08 — Empty CellsCells left empty on *purpose*.\n\nA reference table is only as trustworthy as its blanks. Four values that a reader might expect to find were not located on a vendor page, and rather than fill them from secondary sources the cells say so. A fifth item below is not a missing value at all but a whole row we excluded.\n\n**OpenAI’s supported-language count for GPT-5.5.** “58 languages” and “95+ languages” both circulate on third-party trackers; no OpenAI page stating either was found. Row 14 reads “not located.”\n\n**Google’s language count for Gemini 3.1 Pro.** The Gemini API documentation describes Gemini 3.5 Live Translate, the real-time speech product, as supporting “70+ languages.” That is a different product from the Pro model Tencent benchmarked, and the figure is not carried into row 15. Our [guide to Gemini Live 3.5 Translate](/blog/gemini-live-3-5-translate-real-time-multilingual-cx-guide) covers the real-time product and its consumer competitors; this table is its batch-translation counterpart.\n\n**DeepL Developer-tier overage.** The rendered plans page shows the free allowance and no overage rate; the “$5 per million characters” figure exists only on aggregator sites in this pass.\n\n**DeepL legacy API Pro discontinuation.** Reported by third-party pricing trackers for July 2026; not confirmed on a DeepL page. Carried as reported.\n\n**ALMA — an exclusion, not a blank.** Left out of the table entirely because its maintenance status as of August 2026 could not be confirmed in the time available. An unverified row is worse than a missing one.\n\nAnthropic is the useful contrast. Its [multilingual-support documentation](https://platform.claude.com/docs/en/build-with-claude/multilingual-support) publishes no count at all, by design: “Claude processes input and generates output in most world languages that use standard Unicode characters. Performance varies by language, with particularly strong capabilities in widely spoken languages.” It then benchmarks 14 languages plus English as a percentage of English performance. That is a qualitative framing with a measured floor, and it is the honest shape of a frontier-LLM language claim — which is why the frontier rows in this table carry no number.\n\n## 09 — Reading the TableFour questions the columns *actually* answer.\n\nThe table does not rank systems, because the rows are not competing for one job. It does answer four narrower questions a buyer tends to ask, and the matrix below maps each to the columns that carry the evidence.\n\n##### Can I ship this commercially, self-hosted?\n\nColumn six. Apache 2.0 on Hy-MT2 (all sizes) and MADLAD400-3B-MT; CC-BY-NC on NLLB-200-3.3B, SeamlessM4T v2 Large and TowerInstruct-Mistral-7B-v0.2 — non-commercial by their model cards. Licence read from file on two rows, from the card on the rest.\n\n##### Sentences, or *documents*?\n\nColumn five. NLLB caps at 512 tokens with a degradation warning; Hy-MT2 is served at 8,192 on OpenRouter and allows 262,144 by config. Frontier LLMs are not compared on context here but are built for long inputs; their translation quality at length is not measured in this dataset.\n\n##### Does it cover my language pair?\n\nColumn four. Counts range from 10 (TowerInstruct) to 200 (NLLB) to 400+ trained (MADLAD). Hy-MT2 lists 36 ISO codes including four minority languages. Commercial APIs state coverage loosely; frontier vendors state it qualitatively or not at all.\n\n##### Who measured the score I am reading?\n\nColumn eight. Every Hy-MT2 and frontier-LLM score is one Tencent run. No commercial API publishes a score of its own; the single Microsoft-Translator figure in row 11 is also Tencent's run. Specialist rows 4–7 cite their own papers without re-extracted numbers. Treat the table as a map of what exists, not a leaderboard.\n\nWho needs this table at all: teams translating product catalogues, support content and storefronts into more than a handful of languages, where per-character pricing compounds quickly and where a non-commercial licence on an otherwise excellent model is a blocker. Our [cross-border ecommerce guide](/blog/cross-border-ecommerce-international-selling-guide-2026) describes that buyer; this table supplies the inputs they would otherwise assemble by hand. Teams deciding whether a translation specialist, a commercial API or a prompted frontier model fits a specific pipeline can scope that evaluation through our [AI transformation services](/services/ai-transformation), which begin with exactly this kind of side-by-side.\n\n## 10 — ConclusionA table that keeps *both* columns.\n\n### Specialists and frontier LLMs win different metrics, and the unit matters more than the headline.\n\nThe dataset supports a narrow claim and resists a broad one. The narrow claim: in Tencent’s own evaluation, a 30B-A3B translation specialist under Apache 2.0 posts the best WMT25 GEMBA score of every compared system, and the two largest frontier LLMs take the reference-based XCOMET-XXL column on FLORES-200 EN⇔XX, a higher-scoring test set on which everything lands in the nineties. The broad claim — that specialists have caught the frontier, or that frontier LLMs have made specialists redundant — is supported by neither column.\n\nThe more durable finding is structural. The three categories are priced in different units, licensed on different terms, described with different precision about language coverage, and evaluated by different parties or not at all. A buyer comparing them is comparing a per-token listing against a per-character list price, a LICENSE.txt against a terms page, a count of ISO codes against “most world languages.” Making those differences *visible* rather than collapsing them into one score is the purpose of the table.\n\nLooking ahead, the cells most likely to change are the ones least settled now: an independent reproduction of Table 2 would turn Tencent-run scores into verified ones; a vendor-published language count from OpenAI or Google would fill two blanks; and the gap between OpenRouter’s 8,192-token serving cap and the model’s 262,144-token ceiling invites a long-document evaluation we could not find published anywhere. When any of those land, this page will be refreshed in place and re-dated.", "url": "https://wpnews.pro/news/translation-specialists-vs-frontier-llms-a-reference-table", "canonical_source": "https://www.digitalapplied.com/blog/specialist-translation-models-vs-frontier-llms-reference", "published_at": "2026-08-22 00:00:00+00:00", "updated_at": "2026-08-23 16:14:04.015532+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "large-language-models", "ai-research", "ai-products"], "entities": ["Tencent", "Hy-MT2", "OpenRouter", "GPT-5.5", "Gemini 3.1 Pro", "NLLB-200-3.3B", "SeamlessM4T v2 Large", "TowerInstruct-Mistral-7B-v0.2"], "alternates": {"html": "https://wpnews.pro/news/translation-specialists-vs-frontier-llms-a-reference-table", "markdown": "https://wpnews.pro/news/translation-specialists-vs-frontier-llms-a-reference-table.md", "text": "https://wpnews.pro/news/translation-specialists-vs-frontier-llms-a-reference-table.txt", "jsonld": "https://wpnews.pro/news/translation-specialists-vs-frontier-llms-a-reference-table.jsonld"}}