Best Translation Models for English-Spanish in 2026: Hy-MT2-1.8B and Hunyuan-MT-7B

Quick Answer

Hy-MT2-1.8B (tencent) is the top pick for local English–Spanish translation as of September 2026 because it has the newest release date among the candidates, with no public benchmark yet to establish a translation-quality winner [3]. In order, the ranking is Hy-MT2-1.8B (tencent), Hunyuan-MT-7B (tencent), Seed-X-PPO-7B (ByteDance-Seed), Seed-X-Instruct-7B (ByteDance-Seed), madlad400-3b-mt (google), madlad400-10b-mt (google), nllb-200-distilled-600M (facebook), and nllb-200-distilled-1.3B (facebook) [3][5].

Key Takeaways

• Tencent’s Hy-MT2-1.8B ranks first because its May 11, 2026 release is the newest in this ranking; its BF16 weights occupy 4.1 GB, and its license is Apache-2.0. [3]
• Ranking note: no public benchmark yet scores these candidates against each other, so the ordering is newest release first, then downloads. Too few releases fall within the requested twelve-month window, so the newest available releases are included. [1][2][3][5][6][8][10][11] Eligibility is limited to translation-tagged models from labs with a published paper or leaderboard record, or other publishers exceeding 10,000 downloads in the last thirty days. [1][2][4][7][9]
• Established picks bypassing the recency gate are Google’s madlad400-3b-mt and Facebook’s nllb-200-distilled-600M and nllb-200-distilled-1.3B, the highest-download candidates here; the exemption does not confer first place. Google’s model uses Apache-2.0, while both Facebook models use CC-BY-NC-4.0. [1][2][8]
• Tencent’s Hunyuan-MT-7B has 16.1 GB of BF16 weights and a 32K-token context window; the weight size alone should not be treated as a total GPU-memory requirement. [5]
• ByteDance-Seed’s Seed-X-PPO-7B and Seed-X-Instruct-7B share a July 16, 2025 publication date, 15.0 GB of BF16 weights and the openmdw license. PPO precedes Instruct on downloads—1,280 versus 1,170—not a demonstrated English–Spanish quality advantage. [6][11]
• Google’s madlad400-10b-mt has 42.9 GB of F32 weights, compared with 11.8 GB for madlad400-3b-mt; both use Apache-2.0. Those figures describe published weights, not measured runtime memory or translation quality. [8][10]

How do local English–Spanish translation models compare on specifications and benchmarks?

Which translation model should you run locally for English–Spanish?

1. Hy-MT2-1.8B

Hy-MT2-1.8B by tencent ranks first under the release-recency rule, following its Hugging Face publication on May 11, 2026.[3] The ordering is newest release first, then downloads; no public benchmark yet establishes its English–Spanish quality against the other candidates. Use it as a starting point for local translation evaluation, with your own English–Spanish documents deciding whether to adopt it.

The model reports 2B parameters, a 256K-token context window and 4.1 GB of BF16 weights.[3] Its Apache-2.0 license makes it a candidate for projects that require permissively licensed weights.[3] Document translation is a practical use to evaluate given that context window, but context capacity alone does not establish translation accuracy.

For local hardware planning, quantized weights would occupy approximately 1 GB at 4-bit precision, estimated (params x 0.5 bytes) from the reported 2B parameters.[3] Treat that estimate as a weight-storage budget, not a total GPU or system RAM requirement; runtime memory remains an unverified deployment constraint.

2. Hunyuan-MT-7B

Hunyuan-MT-7B by tencent ranks second under the newest-release-first ordering, behind tencent’s Hy-MT2-1.8B.[3][5] Its Hugging Face publication date is August 28, 2025, placing it outside the annual release window; it remains included because the eligible field has too few recent releases.[5] Hunyuan-MT-7B has no public benchmark yet, so its position does not establish English–Spanish translation quality.

The model has 8B parameters, a 32K-token context window and 16.1 GB of BF16 weights.[5] For local hardware planning, those weights alone do not establish a complete RAM or VRAM requirement. Quantized weight storage would be approximately 4 GB at 4-bit precision—estimated (params x 0.5 bytes), using the cited 8B parameter count.[5] Treat that estimate as a weight budget, not a guarantee that the model fits your hardware.

Use Hunyuan-MT-7B as a candidate for local English–Spanish evaluation on representative documents. Compare terminology, omissions and meaning preservation before adopting it for unattended translation.

3. Seed-X-PPO-7B

Seed-X-PPO-7B by ByteDance-Seed was published on Hugging Face on July 16, 2025.[6] Its position follows “newest release first, then downloads”: the Tencent releases precede it by publication date,[3][5] while its 1,280 monthly downloads place it ahead of the same-day Seed-X-Instruct-7B by ByteDance-Seed, with 1,170.[6][11] The benchmark status is no public benchmark yet; placement does not establish superior English–Spanish translation quality.

The model has 7.5 billion parameters, a 32K-token context window and an openmdw license.[6] Published BF16 weights occupy 15.0 GB.[6] Quantized weight storage would be approximately 3.75 GB at four-bit precision—estimated (params x 0.5 bytes), using the cited 7.5-billion parameter count.[6] Treat those figures as weight-storage budgets, not complete GPU-memory or system-RAM requirements.

Consider it for local English–Spanish evaluation when you can compare translations against your own terminology and reference text. Check the license against your intended deployment. The release falls outside the requested twelve-month window; it appears here because the shortlist includes older available releases.[6]

4. Seed-X-Instruct-7B

Seed-X-Instruct-7B by ByteDance-Seed ranks fourth under the release-date-then-downloads ordering.[11][6] Both Seed-X variants were published on July 16, 2025; Instruct follows PPO because its trailing-month downloads were 1,170 versus 1,280 as of September 25, 2026.[11][6] Its English–Spanish translation quality remains unranked: no public benchmark yet.

The model has 7.5 billion parameters, a 32K-token context window, and 15.0 GB of BF16 weights under the openmdw license.[11] For local hardware planning, allow memory for those weights plus inference overhead. Quantized weight storage would be approximately 3.75 GB, estimated (params x 0.5 bytes) from the cited parameter count; that estimate does not establish total RAM or GPU memory requirements.[11]

Use it as a candidate for evaluating local English–Spanish translation on your own documents. The caveat is the missing public benchmark: its position reflects publication date and download activity, so validate translation quality before choosing it for production.

5. madlad400-3b-mt

madlad400-3b-mt from google ranks fifth under the newest-release-first, then-downloads ordering.[8][10] Published on November 27, 2023, it shares its release date with google’s madlad400-10b-mt; its 53,812 downloads put it ahead of that model’s 4,818 in the September 25, 2026 snapshot.[8][10] Its position does not establish superior English–Spanish translation quality: no public benchmark yet.

The model has 2.9B parameters, with F32 weights totaling 11.8 GB.[8] For local hardware planning, that download size is not a complete runtime memory requirement. A quantized weight budget would be 1.45 GB, estimated (params x 0.5 bytes) from the cited 2.9B parameters.[8] Treat that estimate as weight storage alone, not a verified RAM or GPU fit.

Consider it for a local English–Spanish workflow where Apache-2.0 licensing is useful.[8] The practical caveat is unverified task quality: evaluate representative sentences and documents before adopting it, particularly where terminology and meaning must remain consistent.

6. madlad400-10b-mt

madlad400-10b-mt by google ranks sixth under “newest release first, then downloads”: its publication date matches google’s madlad400-3b-mt, but its download count is lower.[8][10] Both appeared on November 27, 2023; their monthly downloads were 4,818 and 53,812, respectively, as of September 25, 2026.[8][10] The older releases remain because the eligible pool has too few recent candidates.

The model has 10.7 billion parameters, F32 weights totaling 42.9 GB, and an Apache-2.0 license.[10] For local hardware planning, a hypothetical quantized weight footprint is 5.35 GB at 4-bit precision, estimated (params x 0.5 bytes) from the cited parameter count.[10] Budget additional RAM or VRAM for execution; that weight-only estimate does not establish a specific GPU fit.

Consider it for local English–Spanish evaluation when Apache-2.0 licensing suits your deployment requirements.[10] The caveat is translation quality: no public benchmark yet establishes its English–Spanish standing against these candidates, so its position reflects release date and downloads rather than demonstrated accuracy.

7. nllb-200-distilled-600M

nllb-200-distilled-600M from facebook occupies position seven as an established pick, with a Hugging Face publication date of July 8, 2022.[1] The ordering uses newest release first, then downloads. Its 855,123 downloads over the preceding month place it ahead of the same-date facebook nllb-200-distilled-1.3B, with 306,368 downloads.[1][2] English–Spanish quality remains unranked: no public benchmark yet scores this candidate against the other candidates.

The context limit is 1K tokens, making short passages a sensible starting point for evaluation.[1] Specific GPU memory and system RAM requirements are unconfirmed. Before committing hardware, run representative passages locally and measure memory use with your chosen runtime; the model name alone does not establish a deployment memory budget.

Consider it for noncommercial English–Spanish translation experiments involving short text. The caveat is licensing: cc-by-nc-4.0 restricts commercial use, so a commercial deployment needs a different licensing path or model.[1]

8. nllb-200-distilled-1.3B

nllb-200-distilled-1.3B by facebook ranks eighth under the newest-release-first, then-downloads ordering: it was first published on Hugging Face on 2022-07-08 and recorded 306,368 downloads over the last 30 days as of 2026-09-26.[2] Its release date matches facebook’s nllb-200-distilled-600M, which has more downloads.[1][2] Its benchmark status for this English–Spanish comparison is “no public benchmark yet”; the ordering does not establish a translation-quality advantage.

The context limit is 1K tokens, and the license is cc-by-nc-4.0.[2] Consider it for noncommercial English–Spanish evaluation on short passages. Check terminology and meaning against reference translations from your intended workload before adopting it.

For local deployment, a verified RAM or VRAM requirement is unavailable. Measure memory consumption with your chosen runtime and input lengths before selecting hardware. The practical caveat is licensing: the noncommercial restriction makes this candidate unsuitable for a commercial deployment without separate permission.[2]

What hardware do you need to run English–Spanish translation models locally?

Choose hardware around the memory required for your chosen weight format: Tencent’s Hy-MT2-1.8B provides a starting point at 4.1 GB of BF16 weights. [3] Treat weight size as a budgeting baseline, not a guarantee that a device with that much memory can run the model.

Tencent’s Hunyuan-MT-7B has 16.1 GB of BF16 weights. [5] ByteDance-Seed’s Seed-X-PPO-7B and Seed-X-Instruct-7B each have 15.0 GB of BF16 weights. [6][11] Google’s madlad400-3b-mt has 11.8 GB of F32 weights, while Google’s madlad400-10b-mt has 42.9 GB of F32 weights. [8][10] Check the weight format before comparing download sizes or planning system RAM and GPU VRAM.

Quantization offers a different planning baseline. Using the listed parameter counts, weight-only memory would be 1.0 GB for Hy-MT2-1.8B, estimated (2B params x 0.5 bytes), and 3.75 GB for either Seed-X variant, estimated (7.5B params x 0.5 bytes). [3][6][11] Those calculations describe hypothetical quantized weights; they do not establish that a compatible checkpoint exists or that a particular GPU can run it.

Before buying hardware, verify the intended checkpoint and runtime’s total memory requirements. Allow headroom beyond weights, and measure translation latency with your intended English–Spanish workload rather than inferring speed from model size.

Which translation model licenses allow commercial use?

Apache-2.0 and OpenMDW licenses allow commercial use: the Apache-2.0 options are Tencent’s Hy-MT2-1.8B [3] and Google’s madlad400-3b-mt [8] and madlad400-10b-mt [10]; the OpenMDW options are ByteDance-Seed’s Seed-X-PPO-7B [6] and Seed-X-Instruct-7B [11].

For a commercial English–Spanish deployment, those models form the license-compatible shortlist. Tencent’s Hy-MT2-1.8B uses Apache-2.0 [3], as do both Google MADLAD checkpoints listed here [8][10]. ByteDance-Seed’s checkpoints offer a separate licensing route under OpenMDW [6][11]. Review the applicable license before distributing weights, shipping a modified checkpoint or incorporating a model into a customer-facing product.

Facebook’s nllb-200-distilled-600M [1] and nllb-200-distilled-1.3B [2] use CC-BY-NC-4.0, which restricts use to noncommercial purposes. Neither checkpoint belongs on a commercial shortlist under that published license [1][2]. Downloadable weights and local execution do not remove the noncommercial restriction [1][2].

Tencent’s Hunyuan-MT-7B needs a separate license check before commercial adoption: its license is not specified in the information available here [5]. Treat commercial permission as unconfirmed until you have checked the repository’s applicable terms. For deployment planning, keep that unresolved status separate from the explicitly identified Apache-2.0 and OpenMDW options.

How should you evaluate English–Spanish translation quality when there is no public benchmark yet?

Evaluate English–Spanish translation quality with a blinded comparison on your own material, using bilingual reviewers and a written error rubric before choosing a model.

Build an evaluation set that reflects your workload: documentation, support messages, conversational text and domain terminology. Evaluate English-to-Spanish and Spanish-to-English separately. Specify the intended Spanish locale and audience, and include passages where surrounding context determines meaning, tone or pronoun choice.

Compare candidates such as Tencent’s Hy-MT2-1.8B [3] and ByteDance-Seed’s Seed-X-PPO-7B [6] on identical passages. Record each model’s prompt, decoding settings, quantization and software version. Keep development examples separate from the evaluation set so prompt adjustments do not turn into tuning against the evaluation answers.

Hide model names and shuffle output order. Ask reviewers to check meaning preservation, omissions, additions, negation, terminology, grammatical agreement and register. Mark errors that change an instruction or factual meaning separately from stylistic preferences. Allow valid alternative translations instead of requiring an exact match to a reference.

Report results by direction, locale and document type, alongside representative errors and reviewer disagreements. Date the evaluation and describe the workload it covers. Measure latency and memory separately from translation quality. Define acceptance criteria around errors your application can tolerate, and require human review for outputs that fail those criteria.

Frequently Asked Questions

Which model should I try first for local English–Spanish translation?

Tencent’s Hy-MT2-1.8B takes first place because its publication date makes it the newest eligible release, rather than because of a demonstrated English–Spanish benchmark advantage.[3][5][6][8][10][11][1][2] Its status is “no public benchmark yet.” The model has Apache-2.0 licensing, a listed context length of 256K tokens and BF16 weights totaling 4.1 GB.[3] Treat the recommendation as a starting point for evaluating your own English–Spanish documents.

What is the ranking order, and how was it decided?

Ranking note: no public benchmark scores any candidate, so the ordering is newest release first, then downloads. The order is Tencent’s Hy-MT2-1.8B,[3] Tencent’s Hunyuan-MT-7B,[5] ByteDance-Seed’s Seed-X-PPO-7B,[6] ByteDance-Seed’s Seed-X-Instruct-7B,[11] Google’s madlad400-3b-mt,[8] Google’s madlad400-10b-mt,[10] Facebook’s nllb-200-distilled-600M,[1] then Facebook’s nllb-200-distilled-1.3B.[2] Every entry carries the status “no public benchmark yet”; the positions do not establish comparative English–Spanish translation quality. Downloads break release-date ties.

Why does the ranking include older translation models?

Ranking note: scope admits only translation-tagged models from labs with a published paper or leaderboard record, or other publishers above 10,000 Hugging Face downloads in the last 30 days.[1][2][3][4][7][9] Too few recent releases qualify, so the newest available are included.[3][5][6][8][10][11][1][2] Established picks—the three most-downloaded candidates, nllb-200-distilled-600M, nllb-200-distilled-1.3B and madlad400-3b-mt—bypass the recency gate, follow the same ordering rule and cannot take first place through that exemption.[1][2][8]

How much memory should I budget for local translation?

Start with the published weight sizes: Hy-MT2-1.8B has 4.1 GB of BF16 weights,[3] Hunyuan-MT-7B has 16.1 GB,[5] and both Seed-X variants have 15.0 GB each.[6][11] Google lists F32 weights of 11.8 GB for madlad400-3b-mt and 42.9 GB for madlad400-10b-mt.[8][10] Weight sizes alone do not establish total runtime memory requirements. Verify your chosen runtime and workload before selecting hardware.

Which licenses should I check before commercial deployment?

Hy-MT2-1.8B and both listed MADLAD models use Apache-2.0.[3][8][10] Both Seed-X variants list openmdw.[6][11] Both NLLB variants list cc-by-nc-4.0, whose noncommercial restriction matters for commercial deployment.[1][2] Hunyuan-MT-7B’s license is unspecified here; verify its terms before deployment.[5] Choose according to your intended use and review the applicable license, rather than treating downloadable weights as blanket permission.

Should I choose Seed-X PPO or Seed-X Instruct?

Neither variant has a public benchmark here that establishes an English–Spanish quality advantage: both carry “no public benchmark yet.” Both were published on July 16, 2025, with 7.5B parameters, 32K-token context, 15.0 GB of BF16 weights and openmdw licensing.[6][11] PPO ranks ahead only through the downloads tiebreak: 1,280 versus 1,170 over the preceding 30 days, as of September 25, 2026.[6][11]

Sources

1. facebook/nllb-200-distilled-600M model card (Hugging Face) — 2026-09-25
2. facebook/nllb-200-distilled-1.3B model card (Hugging Face) — 2026-09-25
3. tencent/Hy-MT2-1.8B model card (Hugging Face) — 2026-09-25
4. Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild — 2026-05-21
5. tencent/Hunyuan-MT-7B model card (Hugging Face) — 2026-09-25
6. ByteDance-Seed/Seed-X-PPO-7B model card (Hugging Face) — 2026-09-25
7. Seed-X: Building Strong Multilingual Translation LLM with 7B Parameters — 2025-07-18
8. google/madlad400-3b-mt model card (Hugging Face) — 2026-09-25
9. MADLAD-400: A Multilingual And Document-Level Large Audited Dataset — 2023-09-09
10. google/madlad400-10b-mt model card (Hugging Face) — 2026-09-25
11. ByteDance-Seed/Seed-X-Instruct-7B model card (Hugging Face) — 2026-09-25