Best Translation Models to Run Locally in 2026

Quick Answer

Hy-MT2-1.8B (tencent) is the top pick as of September 2026 because it is the newest eligible release; the ranking admits only translation models with the Hugging Face pipeline tag translation, includes the newest available family releases because too few qualify within the last 12 months, and follows newest release first, then downloads, with no public benchmark yet for any candidate [2] [12] [4] [6] [11] [8] [10] [1] [5] [14]. In order, the ranking is Hy-MT2-1.8B (tencent), LFM2-350M-ENJP-MT (LiquidAI), Hunyuan-MT-7B (tencent), Seed-X-PPO-7B (ByteDance-Seed), Seed-X-Instruct-7B (ByteDance-Seed), madlad400-3b-mt (google), madlad400-10b-mt (google), and nllb-200-distilled-600M (facebook) [2][12].

Key Takeaways

• Ranking note: only translation models with Hugging Face’s translation pipeline tag qualify; ordering is newest release first, then downloads. Benchmark status: no public benchmark yet. Too few candidates fall within the past twelve months, so the newest available releases are included despite older publication dates. [2][12][4][6][11][8][10][1]
• Tencent’s Hy-MT2-1.8B ranks first because its Hugging Face publication date, 2026-05-11, is the newest in this ranking. [2][12] Listed specifications are 2B parameters, 256K tokens of context and Apache-2.0 licensing; weight memory at four-bit precision is approximately 1 GB, estimated (params x 0.5 bytes). [2]
• LiquidAI’s LFM2-350M-ENJP-MT has 354M parameters, corresponding to approximately 177 MB of weight memory at four-bit precision, estimated (params x 0.5 bytes). Its listed context is 128K tokens, and its license is lfm1.0. [12]
• Tencent’s Hunyuan-MT-7B is listed at 8B parameters, corresponding to approximately 4 GB of weight memory at four-bit precision, estimated (params x 0.5 bytes), with 32K tokens of context. [4]
• ByteDance-Seed’s Seed-X-PPO-7B and Seed-X-Instruct-7B each have 7.5B parameters: approximately 3.75 GB of weight memory at four-bit precision, estimated (params x 0.5 bytes). Both list 32K tokens of context and the openmdw license [6][11]. Their shared publication date leaves downloads as the ordering tiebreak. [6][11]
• Google’s madlad400-3b-mt and madlad400-10b-mt carry Apache-2.0 licenses. [8][10] Facebook’s nllb-200-distilled-600M carries cc-by-nc-4.0, which restricts commercial use, and lists 1K tokens of context. [1] Weight-memory estimates describe parameter storage, not total runtime memory.

How do local translation models compare on specifications and benchmarks?

Which translation models should you run locally?

1. Hy-MT2-1.8B

Hy-MT2-1.8B by tencent ranks first because its Hugging Face publication date, 2026-05-11, makes it the newest listed candidate.[2] The ranking admits only translation models with Hugging Face’s translation pipeline tag. The ordering is newest release first, then downloads. Too few recent releases qualify, so the newest available family releases are included.

The model card lists 2B parameters, a 256K-token context window, BF16 weights occupying 4.1 GB, and an Apache-2.0 license.[2] For local hardware planning, weight memory at 4-bit is approximately 1 GB, estimated (params x 0.5 bytes) from the cited 2B parameter count.[2] Treat that estimate as a weight-only budget: a machine needs additional memory for inference, so it does not establish total RAM or VRAM requirements.

Consider Hy-MT2-1.8B for local translation workflows that need room for long inputs, given its advertised context window.[2] The caveat is translation quality: no public benchmark yet establishes its position against the other candidates. Its placement reflects release recency rather than demonstrated benchmark superiority.

2. LFM2-350M-ENJP-MT

LFM2-350M-ENJP-MT by LiquidAI ranks second under the newest release first, then downloads rule: its Hugging Face publication date is September 3, 2025,[12] behind Tencent’s Hy-MT2-1.8B, published May 11, 2026.[2] Placement reflects release timing; the benchmark status is “no public benchmark yet.”

The model has 354 million parameters, a 128K-token context window and 0.7 GB of BF16 weights.[12] Quantized weight storage would be approximately 177 MB, estimated (params x 0.5 bytes) from its 354 million parameters.[12] Treat that estimate as a weight-storage budget when choosing local hardware. A supported quantized build, total runtime RAM requirement and tested device configuration are not specified, so the estimate alone cannot establish whether a particular machine can run it.

Consider LiquidAI’s model for local translation experiments where weight storage is a constraint. Deployment carries a licensing caveat: the model uses the lfm1.0 license; check its terms against your intended use.[12]

3. Hunyuan-MT-7B

Hunyuan-MT-7B by tencent ranks third under the release-date ordering, with a Hugging Face publication date of August 28, 2025.[4] The ordering basis is newest release first, then downloads; no public benchmark yet establishes its position on translation quality. The ranking admits only translation models with the Hugging Face pipeline tag translation. Older releases remain eligible because the recent-release pool is too small.

The repository lists 8B parameters, a 32K-token context window and BF16 weights totaling 16.1 GB.[4] For local hardware planning, weight memory at 4-bit is approximately 4 GB, estimated (params x 0.5 bytes) from the cited 8B parameter count.[4] Treat that calculation as a weight-only estimate, not a measured RAM or VRAM requirement.

Consider Hunyuan-MT-7B for local translation workloads where its listed 32K-token context window is useful.[4] Hardware selection still requires checking runtime memory use at your intended input length. The practical caveat is the absence of a public benchmark comparison: its ranking reflects release order, so evaluate translation quality on your own representative text before deployment.

4. Seed-X-PPO-7B

Seed-X-PPO-7B by ByteDance-Seed ranks here under “newest release first, then downloads” among models tagged for translation. Its first Hugging Face publication was July 16, 2025.[6] It precedes ByteDance-Seed’s Seed-X-Instruct-7B on the download tiebreak: 1,280 versus 1,170 downloads over the preceding 30 days as of September 25, 2026.[6][11] Benchmark status: no public benchmark yet; placement does not establish superior translation quality.

The model has 7.5B parameters, a 32K-token context window, and an openmdw license.[6] Published BF16 weights occupy 15.0 GB.[6] For local hardware planning, weight storage at 4-bit is 3.75 GB, estimated (params x 0.5 bytes) from the cited 7.5B parameter count.[6] Allow additional memory for inference; the weight estimate is not a total RAM or VRAM requirement.

Consider Seed-X-PPO-7B for local translation evaluation when its memory footprint fits your hardware. The practical caveat is that download popularity provides no translation-quality guarantee: evaluate representative text before choosing it for production.

5. Seed-X-Instruct-7B

Seed-X-Instruct-7B by ByteDance-Seed ranks here under the newest release first, then downloads ordering. Published on Hugging Face on July 16, 2025, it shares its release date with Seed-X-PPO-7B by ByteDance-Seed.[11][6] The download tiebreak puts it behind that sibling: 1,170 versus 1,280 downloads over the preceding 30 days as of September 25, 2026.[11][6] Its specifications include 7.5 billion parameters, a 32K-token context window, and the openmdw license.[11]

For local hardware planning, its 7.5 billion parameters imply a weight footprint of 3.75 GB at 4-bit, estimated (params x 0.5 bytes); published BF16 weights occupy 15.0 GB.[11] Treat the quantized estimate as a weight-only budget, not a complete RAM or VRAM requirement. Consider it for local translation workflows where the stated context window fits your input and you can evaluate output quality before deployment. The caveat is straightforward: no public benchmark yet, so its position does not establish a translation-quality advantage.

6. madlad400-3b-mt

madlad400-3b-mt by google ranks here under the ordering rule “newest release first, then downloads.” Its Hugging Face publication date is 2023-11-27.[8] For that shared publication date, its 53,812 downloads exceed madlad400-10b-mt by google’s 4,818 downloads over the last 30 days, as of 2026-09-25.[8][10] The placement reflects release timing and a download tiebreak, rather than demonstrated translation quality: no public benchmark yet.

The model has 2.9B parameters; its 4-bit weight footprint is approximately 1.45 GB, estimated (params x 0.5 bytes).[8] Published F32 weights occupy 11.8 GB.[8] For local hardware planning, use the footprint of your chosen weight format as a starting point and budget additional memory for execution. The estimate is a weights-only calculation, not a verified RAM or VRAM requirement.

Consider madlad400-3b-mt for local translation projects where Apache-2.0 licensing is a selection requirement.[8] A practical caveat is the absence of a confirmed context limit or dated benchmark result here; evaluate your intended text lengths and language pairs before deployment.

7. madlad400-10b-mt

madlad400-10b-mt from google ranks behind google’s madlad400-3b-mt under the ordering rule: newest release first, then downloads. Both were first published on Hugging Face on 2023-11-27 [8][10], but their monthly downloads were 4,818 and 53,812 respectively as of 2026-09-25 [8][10]. The older release remains included because the translation-model category has too few recent candidates. Benchmark status: no public benchmark yet.

The model has 10.7B parameters and ships with 42.9 GB of F32 weights under Apache-2.0 [10]. For local hardware planning, its 10.7B parameters imply 5.35 GB of 4-bit weight storage, estimated (params x 0.5 bytes) [10]. Treat that calculation as a weight-storage estimate, not a verified RAM or VRAM requirement.

Consider it for local translation deployments where Apache-2.0 licensing fits your requirements [10]. The practical caveat is hardware sizing: the weight estimate alone does not establish whether your intended runtime and workload will fit. Validate that combination before committing hardware.

8. nllb-200-distilled-600M

nllb-200-distilled-600M by facebook ranks eighth under the ordering “newest release first, then downloads.”[1] Its Hugging Face publication date is 2022-07-08, so it appears as an older fallback in a field with too few recent releases.[1] Its 855,123 downloads over the reporting month break a publication-date tie; popularity does not establish translation quality.[1] Benchmark status: no public benchmark yet.

The context limit is 1K tokens, making short-text translation evaluation a practical use to consider.[1] A verified parameter count, weight size and hardware minimum are unavailable here, so a defensible memory estimate for quantized weights cannot be given. Check the intended runtime’s requirements before choosing a local machine.

The main caveat is the cc-by-nc-4.0 license, which restricts commercial use.[1] Treat the model as a candidate for noncommercial experiments, and evaluate translation quality on representative text before adopting it.

How much memory should you budget for local translation models?

Budget above the weight-only estimate: for tencent’s Hy-MT2-1.8B (tencent), that means more than 1 GB at 4-bit, estimated (params x 0.5 bytes) from its listed 2B parameters.[2] Treat that calculation as a starting point, not a measured total RAM or VRAM requirement.

LiquidAI’s LFM2-350M-ENJP-MT (LiquidAI) has 354M parameters, giving a 4-bit weight budget of 177 MB, estimated (params x 0.5 bytes).[12] google-t5’s google-t5/t5-small has 61M parameters, giving 30.5 MB, estimated (params x 0.5 bytes).[14]

For ByteDance-Seed’s Seed-X-PPO-7B (ByteDance-Seed) and Seed-X-Instruct-7B (ByteDance-Seed), each listed at 7.5B parameters, allow 3.75 GB for quantized weights, estimated (params x 0.5 bytes).[6][11] tencent’s Hunyuan-MT-7B (tencent) is listed at 8B parameters, giving 4 GB, estimated (params x 0.5 bytes).[4] Use the listed parameter count when calculating memory rather than inferring it from the model name.

google’s madlad400-3b-mt (google) has 2.9B parameters, giving 1.45 GB, estimated (params x 0.5 bytes).[8] google’s madlad400-10b-mt (google) has 10.7B parameters, giving 5.35 GB, estimated (params x 0.5 bytes).[10]

Keep downloaded weight size separate from your quantized estimate. The MADLAD checkpoints list F32 weights of 11.8 GB and 42.9 GB respectively.[8][10] None of these calculations establishes total runtime memory: leave headroom and measure your intended workload before committing hardware.

Which licenses apply to local translation models?

Local translation models in this selection use Apache-2.0, cc-by-nc-4.0, openmdw or lfm1.0, depending on the model.[1][2][6][12] Check the license attached to the exact repository you plan to deploy; a shared translation task does not establish shared licensing terms.

Apache-2.0 applies to Tencent’s Hy-MT2-1.8B,[2] Google’s madlad400-3b-mt,[8] Google’s madlad400-10b-mt,[10] and google-t5’s t5-small.[14] Those entries share a license label, so you can group them for an initial license review while keeping each model’s documentation with its deployment record.

Facebook’s nllb-200-distilled-600M and nllb-200-3.3B carry cc-by-nc-4.0.[1][5] Review that license against your intended use before selecting either model for a project, particularly a commercial deployment.

ByteDance-Seed’s Seed-X-PPO-7B and Seed-X-Instruct-7B both list openmdw.[6][11] LiquidAI’s LFM2-350M-ENJP-MT lists lfm1.0.[12] Read those named agreements directly rather than applying assumptions from the Apache-licensed entries.

For Tencent’s Hunyuan-MT-7B, verify the applicable license in the original repository before deployment.[4] Make license confirmation part of model selection: record the repository, retain the applicable license text, and review its conditions against your planned use and distribution. Treat local execution and permission to use or redistribute a model as separate checks.

How are translation models ranked when no public benchmark scores them?

When no public benchmark scores the eligible translation models, the ordering is newest release first, then downloads, using Hugging Face publication dates and download counts.[2][12][4][6][11][8][10][1][5][14]

The ranking admits only translation models with the Hugging Face pipeline tag translation. The intended release window is the past year, but the category has too few qualifying releases, so the newest available candidates are included with older publication dates stated explicitly.[2][12][4][6][11][8][10][1][5][14]

tencent’s Hy-MT2-1.8B (tencent) ranks first because its Hugging Face publication date is the newest among the admitted candidates: May 11, 2026.[2][12][4][6][11][8][10][1][5][14] Placement reflects release recency; translation quality remains unranked without a comparable public benchmark.

Downloads break ties between candidates published on the same date. ByteDance-Seed’s Seed-X-PPO-7B (ByteDance-Seed) and Seed-X-Instruct-7B (ByteDance-Seed) were both published on July 16, 2025.[6][11] Their respective download counts were 1,280 and 1,170 over the preceding 30 days, as of September 25, 2026, placing the PPO release ahead.[6][11] Popularity serves only as a tiebreaker, and parameter count does not determine placement.

Every candidate carries the benchmark status “no public benchmark yet.” Any model-card benchmark must be labeled “self-reported.” A leaderboard may compare only models evaluated together, and an older scored model cannot displace a newer unscored release under these ranking rules.

Frequently Asked Questions

Which local translation model should I try first?

Hy-MT2-1.8B (tencent) ranks first because its publication date is the newest among the candidates: 2026-05-11.[2] Ranking note: eligibility is limited to translation models carrying Hugging Face’s translation pipeline tag. The family has too few recent releases, so the newest available are included, including older entries.[1][8] The ordering is newest release first, then downloads; no public benchmark scores any candidate.

How do the local translation models rank?

The order is Hy-MT2-1.8B (tencent)[2], LFM2-350M-ENJP-MT (LiquidAI)[12], Hunyuan-MT-7B (tencent)[4], Seed-X-PPO-7B (ByteDance-Seed)[6], Seed-X-Instruct-7B (ByteDance-Seed)[11], madlad400-3b-mt (google)[8], madlad400-10b-mt (google)[10], nllb-200-distilled-600M (facebook)[1], facebook/nllb-200-3.3B[5], and google-t5/t5-small from google-t5.[14] Downloads only break publication-date ties. Parameter count does not determine position, and an older scored model cannot displace a newer unscored model.

How much memory should I budget for quantized weights?

For Hy-MT2-1.8B (tencent), 2B parameters imply 1 GB at 4-bit, estimated (params x 0.5 bytes).[2] For Seed-X-PPO-7B (ByteDance-Seed), 7.5B parameters imply 3.75 GB at 4-bit, estimated (params x 0.5 bytes).[6] For madlad400-10b-mt (google), 10.7B parameters imply 5.35 GB at 4-bit, estimated (params x 0.5 bytes).[10] Treat those figures as weight-storage estimates, not measured total runtime memory or guarantees that inference will fit.

Which compact models should I compare for a local setup?

Compare LFM2-350M-ENJP-MT (LiquidAI), listed at 354M parameters with a 128K-token context, with google-t5/t5-small, listed at 61M parameters.[12][14] LiquidAI’s model was published on 2025-09-03, while google-t5’s model was published on 2022-03-02.[12][14] The ranking therefore places LiquidAI’s model earlier. Neither candidate has a public benchmark here that establishes a translation-quality advantage; evaluate your intended language pair before choosing.

What licenses do these translation models use?

Hy-MT2-1.8B (tencent), madlad400-3b-mt (google), madlad400-10b-mt (google), and google-t5/t5-small list Apache-2.0.[2][8][10][14] Seed-X-PPO-7B (ByteDance-Seed) and Seed-X-Instruct-7B (ByteDance-Seed) list openmdw.[6][11] LFM2-350M-ENJP-MT (LiquidAI) lists lfm1.0.[12] nllb-200-distilled-600M (facebook) and facebook/nllb-200-3.3B list cc-by-nc-4.0.[1][5] Check the applicable license before deployment, especially for a commercial application. A license designation is not a translation-quality score or a hardware compatibility guarantee.

Does the ranking prove which model translates better?

No public benchmark scores these candidates against each other, so every entry carries the status “no public benchmark yet.” Publication dates determine the ranking before download tiebreaks. The Hy-MT2 paper is dated 2026-05-21, and the Seed-X paper is dated 2025-07-18; neither date is a benchmark result.[3][7] Any model-card benchmark must be labeled self-reported rather than treated as an independent leaderboard comparison.

Sources

1. facebook/nllb-200-distilled-600M model card (Hugging Face) — 2026-09-25
2. tencent/Hy-MT2-1.8B model card (Hugging Face) — 2026-09-25
3. Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild — 2026-05-21
4. tencent/Hunyuan-MT-7B model card (Hugging Face) — 2026-09-25
5. facebook/nllb-200-3.3B 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
12. LiquidAI/LFM2-350M-ENJP-MT model card (Hugging Face) — 2026-09-25
13. LFM2 Technical Report — 2025-11-28
14. google-t5/t5-small model card (Hugging Face) — 2026-09-25
15. Neural Network Acceptability Judgments — 2018-05-31