Best open-source LLM for multilingual translation

Find a Open-source translation LLM being performant without relying on a proprietary API has become a concrete priority for teams handling large volumes of multilingual content (contracts, customer support, technical documentation). Unlike a cloud API billed by the token, a locally hosted open-weight model provides complete control over translated data and a fixed cost independent of volume. This article reviews the general-purpose models in the catalog that are best suited to translation, the required hardware configuration, and the role of specialized models such as Aya 32B or Madlad-400 in this landscape.

Why choose an open-weights model for translation

Translating sensitive documents (legal, HR, medical) raises a confidentiality issue when the text passes through a third-party API. A model running locally via llama.cpp, vLLM or Ollama eliminates this network transit: the files remain on the company's infrastructure. That's the reasoning detailed in our guide on document translation with a local LLM and in the one dedicated to multilingual customer support.

The second advantage is the license. Most recent multilingual models in the catalog are published under Apache 2.0 or MIT, allowing commercial use without royalties—a point you should systematically verify before deploying to production.

Multilingual general-purpose models to prioritize

The large mixture-of-experts models released in 2025-2026 generally natively support dozens of languages, thanks to massively multilingual training corpora. Options in the catalog include:

• Qwen 3.5 397B-A17B — 397B parameters, Apache 2.0 license, estimated Q4 VRAM of ~240 GB, 262,000-token context. Alibaba trains its Qwen family on a highly multilingual corpus including many Asian and European language pairs (model card, Alibaba on Hugging Face).
• Mistral Large 3 675B — 675B parameters, Apache 2.0, ~405 GB Q4 VRAM, 256,000-token context. Mistral AI emphasizes European languages in its models from the design stage (model card, Mistral AI on Hugging Face).
• GLM 5.3 Flash 320B-A18B — 320B parameters, MIT license, ~186 GB Q4 VRAM, 128,000-token context, a lighter-to-deploy option than models above 400B (model card, Zhipu AI on Hugging Face).
• DeepSeek V3.2 — 685B parameters, MIT license, ~410 GB VRAM in Q4, 128,000-token context (model card, DeepSeek on Hugging Face).
• Mistral Small 4 — 119B parameters, Apache 2.0, Q4 VRAM ~72 GB, 256,000-token context: the best option in this comparison for a single-GPU 80 GB machine or a unified-memory Mac (model card).

For long text (contracts, technical manuals), prioritizing a large context matters as much as model size: beyond 100,000 tokens, the entire document fits in a single request without manual splitting, preserving terminology consistency.

Specialized translation models: Aya 32B and Madlad-400

Outside the catalog's general-purpose models, two families are specifically designed for multilingual translation and frequently appear in comparisons: Aya 32B (Cohere's Aya Expanse family, focused on covering more than 20 languages) and Madlad-400 (trained on a corpus covering more than 400 languages). These models are not included in BestLLMfor's main catalog, and their exact specifications (VRAM per quantization, tokens/sec) must be confirmed on a case-by-case basis depending on the variant tested.

To compare Aya 32B with comparable general-purpose models, see our comparison Aya Expanse 32B vs Qwen 2.5 32B as well as the lighter version Aya Expanse 8B vs Llama 3 8B. In practice, a specialized model like Aya excels at languages with limited training data, while a large general-purpose model like Qwen 3.5 or Mistral Large 3 remains more versatile for mixed content (translation + rewriting + summarization in the same pipeline).

Hardware configuration and quantization

The quantization choice directly determines the required VRAM:

• Q4 : this is the reference level used for the figures in this catalog—approximately 0.58 to 0.6 GB per billion parameters for recent MoE architectures.
• Q8 : about twice the Q4 VRAM (estimated), with a generally negligible loss in translation quality.
• FP16 : approximately four times the Q4 VRAM (estimated), reserved for deployments with multiple high-end GPUs.

For a workstation with a single 24 GB or 32 GB GPU, only compact models such as Mistral Small 4 in Q4 quantization remain accessible; models above 200B require a cluster or a Mac with a large amount of unified memory. Our European multilingual guide and our selection of French-language models detail the VRAM/quality tradeoffs for more modest configurations.

Tokens/sec throughput varies significantly with the GPU, context length, and concurrent load; without a measurement on the target hardware, this figure should be confirmed rather than estimated in advance.

Concrete use cases

• Multilingual customer support : a model such as GLM 5.3 Flash, which is lighter to host, can process tickets in multiple languages with controlled latency — details in the multilingual customer support guide.
• Long-document translation : Mistral Large 3 or DeepSeek V3.2, with their extended context, avoid splitting contracts or manuals — see the document translation guide.
• Quality evaluation : before deployment, compare the candidate models’ multilingual scores on theOpen LLM Leaderboard and check the model documentation at Hugging Face Hub.

FAQ

Q: What is the best open-source LLM for high-volume translation?

For a large workload with hardware available, Qwen 3.5 397B-A17B or Mistral Large 3 675B offer good multilingual coverage under an Apache 2.0 license. For more modest infrastructure, Mistral Small 4 (119B, Q4 VRAM ~72 GB) remains the most accessible option in the catalog.

Q: Is Aya 32B better than the general-purpose models in the catalog?

For languages that are underrepresented in general-purpose corpora, Aya 32B may outperform them. For common language pairs (English–French, English–Chinese), general-purpose models such as Qwen 3.5 or GLM 5.3 Flash are often competitive, with a longer context. See the comparison Aya Expanse 32B vs Qwen 2.5 32B.

Q: Is Madlad-400 suitable for enterprise deployment?

Madlad-400 targets very broad language coverage (more than 400 languages), which is useful for niche use cases. Its exact VRAM requirements by quantization must be confirmed for the variant used; it is not included in this comparison's main catalog.

Q: Do you need a dedicated GPU to run a translation LLM locally?

It depends on the size of the model you choose. A compact model such as Mistral Small 4 in Q4 fits on a 24–32 GB GPU or a Mac with sufficient unified memory. Models over 300B require multiple GPUs or a machine with large unified memory.

Q: Which license should you choose for commercial use?

Prefer Apache 2.0 or MIT, which impose neither royalties nor commercial-use restrictions—this applies to Qwen 3.5, Mistral Large 3, GLM 5.3 Flash, and DeepSeek V3.2 in this comparison. Always verify the exact license on the model page before deployment.

Q: How can I measure translation quality before choosing a model?

Test on a representative sample from your domain (contracts, support tickets, documentation) rather than relying solely on generic benchmarks such as MMLU. Also consult theOpen LLM Leaderboard for an initial indication of general ability before job-specific testing.

Conclusion

Choose one Open-source translation LLM depends primarily on the volume to process and the available hardware: general-purpose models such as Qwen 3.5, Mistral Large 3, or GLM 5.3 Flash cover most use cases under permissive licenses, while specialized models such as Aya 32B or Madlad-400 remain relevant for low-resource languages. To refine this choice for your hardware configuration, use the configurator or explore the entire catalog.