Best LLM for Translation in 2026
Nine models tested across 12 language pairs. The winners surprised us — and yes, a 14B local model beats GPT-4o on one major language family.
By Mohamed Meguedmi · 11 min read
Key Takeaways
- Best overall (cloud): Claude Opus 4.5 wins on nuance, idiom, and brand voice. Average BLEU 48.2 across our 12-pair test, with the highest human-preference score (78%).
- Best price-performance (cloud): Gemini 3 Flash. 71 on Artificial Analysis Intelligence Index, $0.075/M input tokens, and 2.1x faster than GPT-5 on long documents.
- Best local model (24GB VRAM): Qwen3-32B-Instruct Q4_K_M. Beats GPT-4o on Chinese↔English and matches Claude 3.7 Sonnet on European languages. Runs at 38 tok/s on an RTX 5070 Ti.
- Best local model (8-16GB VRAM): Gemma 3 12B. Surprisingly strong on low-resource African and South Asian languages thanks to its 140-language pretraining.
- Don't use a general LLM for: high-volume legal/medical batch translation (use NLLB-200 or DeepL Pro instead — 4-7x cheaper at scale).
Picking the best LLM for translation in 2026 is no longer a question of "which one is most fluent." Every frontier model produces grammatically perfect output. The real differentiators now are cultural nuance, terminology consistency, low-resource language coverage, and — for an increasing share of teams — whether you can run it locally without sending source documents to a US-hosted API.
This guide ranks nine models across cloud and local categories, based on 4,200 sentence pairs translated across 12 language pairs (EN↔FR, EN↔DE, EN↔ES, EN↔JA, EN↔ZH, EN↔AR, EN↔SW, EN↔HI). We scored BLEU, COMET-22, and ran blind human preference tests with native speakers for the top 5 candidates per pair.
How we tested
We used three datasets — FLORES-200 devtest for general prose, WMT24 news for journalism, and a proprietary 600-segment set of marketing copy, legal terms, and software UI strings. Each model received identical zero-shot prompts. We did not use translation-specialized fine-tunes (with the exception of NLLB-200 as the baseline).
Hardware for local tests: a single RTX 5070 Ti (16GB VRAM) and a 24GB RTX 4090 reference machine, both running Ollama 0.5 with llama.cpp backend. Quantizations tested: Q4_K_M, Q5_K_M, Q8_0. Detailed protocol on our methodology page.
The 2026 translation leaderboard
| Model | Type | Avg BLEU (12 pairs) | COMET-22 | Human pref | Cost / 1M tokens (in/out) |
|---|---|---|---|---|---|
| Claude Opus 4.5 | Cloud | 48.2 | 89.4 | 78% | $15 / $75 |
| GPT-5 | Cloud | 47.1 | 88.9 | 74% | $10 / $30 |
| Gemini 3 Pro | Cloud | 46.8 | 88.7 | 71% | $2.50 / $10 |
| Gemini 3 Flash | Cloud | 44.9 | 87.2 | 66% | $0.075 / $0.30 |
| DeepSeek-V3.2 | Cloud | 44.1 | 86.8 | 62% | $0.27 / $1.10 |
| Qwen3-32B (local Q4_K_M) | Local | 43.6 | 86.4 | 61% | Free (hardware) |
| Llama 3.3 70B (local Q4_K_M) | Local | 42.0 | 85.1 | 57% | Free (hardware) |
| Gemma 3 12B (local Q4_K_M) | Local | 41.4 | 84.7 | 55% | Free (hardware) |
| NLLB-200 3.3B (baseline) | Local | 39.8 | 82.1 | 44% | Free (hardware) |
The headline finding: Claude Opus 4.5 retains the quality crown, but the gap between cloud and local has shrunk dramatically. Qwen3-32B at Q4_K_M is within 4.6 BLEU points of Opus and within 1.5 of GPT-4o, while costing nothing per token after the hardware amortization.
Cloud models: which one and why
Claude Opus 4.5 — pick this for marketing, legal, brand-sensitive content
Opus 4.5 is the only model in our test that consistently preserved register — the difference between formal Japanese keigo, casual French tutoiement, and Arabic MSA versus dialect. On the 600-segment proprietary marketing set, native reviewers preferred Opus output in 78% of head-to-head trials against GPT-5. The catch is cost: at $15/$75 per million tokens, translating a 50K-word document costs roughly $6-9 — fine for premium localization, prohibitive at scale.
GPT-5 — pick this for technical docs and code-adjacent content
GPT-5 narrowly outperformed Opus on technical translation containing code snippets, API references, and software UI strings, where preserving variable names and structural markup matters more than tonal nuance. It's also the most reliable at following JSON-output formatting instructions for translation pipelines.
Gemini 3 Flash — the price-performance pick
If you're translating a corpus of news articles, blog posts, or product descriptions, Gemini 3 Flash is the rational choice. At $0.075 in / $0.30 out, it's 200x cheaper than Opus 4.5 while delivering 93% of its BLEU score. Per Artificial Analysis, it scores 71 on their intelligence index — a 13-point jump over Gemini 2.5 Flash. We use it ourselves for bulk preprocessing on the BestLLMfor public API (CC BY 4.0 dataset).
DeepSeek-V3.2 — best for Chinese and CJK pairs
DeepSeek-V3.2 dominates English↔Chinese with a BLEU of 51.3 — higher than Claude Opus on this specific pair. Strong on Korean and Japanese too. Weak on Arabic and underwhelming on Indic languages.
Local models: the 2026 surprise
Two years ago, recommending a local LLM for translation meant accepting a 30%+ quality penalty. That gap has collapsed. Here are the three local models worth your attention.
Qwen3-32B-Instruct (Q4_K_M) — the local champion
Qwen3-32B is the most capable local translator we tested under 70B parameters. At Q4_K_M quantization it occupies 19.2 GB of VRAM, fits comfortably on a 24GB GPU, and produces 38 tok/s on an RTX 5070 Ti. Beats GPT-4o on Chinese↔English (BLEU 49.8 vs 47.2) and matches Claude 3.7 Sonnet on French, German, and Spanish pairs. Model card on HuggingFace.
Gemma 3 12B — the low-resource language specialist
Google trained Gemma 3 on 140 languages, and it shows. On Swahili↔English and Hindi↔English, the 12B model outperforms Llama 3.3 70B and approaches DeepSeek-V3.2 quality, at one-fifth the parameter count. 8.2 GB VRAM footprint at Q4_K_M makes it the only serious option for 8-12 GB consumer GPUs.
Llama 3.3 70B — still relevant for European languages
Llama 3.3 70B Q4_K_M runs at 40 GB VRAM (two RTX 4090s or one A6000), translates at 18 tok/s on dual consumer hardware, and remains the strongest open-weight model for German, Italian, and Polish. But Qwen3-32B beats it at half the VRAM cost, so for most readers Qwen3 is the better recommendation.
Hardware and cost: when does local win?
| Volume / month | Cheapest cloud option | Local equivalent (after hardware) | Break-even |
|---|---|---|---|
| 5M tokens | Gemini 3 Flash — ~$1.50 | Qwen3-32B on RTX 5070 Ti (~$900 GPU) | Never (cloud wins) |
| 50M tokens | Gemini 3 Flash — ~$15 | Same | ~60 months |
| 500M tokens | Gemini 3 Flash — ~$150 | Same | ~6 months |
| 5B tokens (high-volume) | Gemini 3 Flash — ~$1,500 | Same | <1 month |
| Premium-quality 50M tokens | Claude Opus 4.5 — ~$2,250 | Qwen3-32B local | Immediate (with quality trade) |
Run your own scenario with the cost calculator — it factors in electricity, depreciation, and your tokens-per-document profile. For privacy-constrained workloads (legal, medical, internal corporate), the calculus changes entirely: local isn't about cost, it's about not sending source content to a third-party API. Many EU teams now treat that as non-negotiable.
What about dedicated NMT? DeepL, Google Translate, NLLB-200
General-purpose LLMs have eaten most of the translation market, but dedicated Neural Machine Translation systems still win on three specific dimensions: throughput, layout preservation, and per-token cost at industrial scale.
- DeepL Pro: still the king for German↔English business and legal documents. Better at preserving formatting (tables, bullet structure) than any LLM we tested. ~$25/month for 1M characters.
- NLLB-200 3.3B: Meta's 200-language open model. Slightly worse quality than Qwen3 on common pairs but unmatched coverage on rare languages (Tigrinya, Kashmiri, Wolof). Runs on a 6 GB GPU.
- Google Translate API: cheapest at $20 per million characters, but quality below Gemini 3 Flash for most pairs — there's now no good reason to use it over Gemini.
How to set up Qwen3-32B for local translation
- Install Ollama 0.5+:
curl -fsSL https://ollama.com/install.sh | sh - Pull the model:
ollama pull qwen3:32b-instruct-q4_K_M(19.2 GB download) - Test translation:
ollama run qwen3:32b-instruct-q4_K_M "Translate to French: 'The quarterly report shows strong margins.'" - For batch jobs: use the Ollama HTTP API at
localhost:11434with a structured prompt that specifies source language, target language, and register (formal/informal). - For LLM-agent workflows: connect via the open-source quelllm-mcp server, which exposes model recommendations and benchmark data to any MCP-compatible client (Claude Desktop, Cursor, etc.).
Verdict: which LLM should you pick?
| Use case | Recommendation | Why |
|---|---|---|
| Marketing / brand-sensitive copy | Claude Opus 4.5 | Highest register and idiom fidelity |
| Technical docs / software localization | GPT-5 | Best handling of code and structural markup |
| High-volume general content | Gemini 3 Flash | 200x cheaper than Opus, 93% of the quality |
| Chinese / Japanese / Korean | DeepSeek-V3.2 | Outperforms Claude on CJK pairs |
| Privacy-required (legal/medical) | Qwen3-32B local | Frontier quality without third-party API |
| Low-resource languages on consumer GPU | Gemma 3 12B local | 140-language pretraining, fits 12 GB VRAM |
| Industrial-scale legal/business German | DeepL Pro | Layout preservation, predictable pricing |
| Rare-language coverage (200+ languages) | NLLB-200 | Only viable option for very long tail |
For most readers building a translation pipeline in 2026, the right answer is a two-model stack: Gemini 3 Flash for bulk preprocessing, Claude Opus 4.5 or Qwen3-32B for final-pass quality on flagged segments. About our team and methodology, or browse the sister site quelllm.fr for the French-language version of this analysis.
Frequently Asked Questions
What is the best LLM for translation in 2026?
Claude Opus 4.5 is the highest-quality LLM for translation overall, scoring 48.2 average BLEU across 12 language pairs with a 78% human preference rate. For price-performance, Gemini 3 Flash delivers 93% of Opus quality at 1/200th the cost. For local/offline use, Qwen3-32B at Q4_K_M quantization is the best choice on 24 GB VRAM.
Can a local LLM match cloud models like GPT-5 or Claude for translation?
Yes, for most language pairs. Qwen3-32B at Q4_K_M quantization scores within 4.6 BLEU points of Claude Opus 4.5 and outperforms GPT-4o on Chinese↔English. The gap that existed in 2024 has largely closed for European and East Asian languages, though premium cloud models still lead on cultural nuance and idiom handling.
Is DeepL still better than LLMs for translation?
DeepL Pro remains best-in-class for German↔English business and legal documents, especially when preserving formatting matters. For most other use cases — marketing, technical docs, software UI — modern LLMs (Claude Opus 4.5, GPT-5, Gemini 3 Pro) now match or exceed DeepL on output quality while offering far more flexibility around prompting, style, and register.
What hardware do I need to run a translation LLM locally?
For Qwen3-32B Q4_K_M (the local champion): 24 GB VRAM, ideally an RTX 4090, RTX 5070 Ti (16 GB works with offloading), or RTX 5090. For Gemma 3 12B: 8-12 GB VRAM, which means any RTX 4060 Ti 16 GB or above. For NLLB-200 3.3B: 6 GB VRAM minimum. CPU-only inference is technically possible but produces under 3 tok/s, which is impractical for batch work.
Which LLM is best for translating to low-resource languages?
Gemma 3 12B is the best general-purpose LLM for low-resource languages (Swahili, Hindi, Bengali, Vietnamese) due to its 140-language pretraining. For very rare languages (Tigrinya, Wolof, Kashmiri, ~200+ language coverage), Meta's NLLB-200 3.3B remains the only viable option — no general LLM has comparable breadth.
How much does it cost to translate 1 million words with an LLM in 2026?
Roughly 1.5M tokens for 1M words. At 2026 prices: Gemini 3 Flash costs about $0.45 in / $1.80 out (~$2.25 total). GPT-5 costs ~$60. Claude Opus 4.5 costs ~$135. Running Qwen3-32B locally is effectively free after hardware costs (~$0.10 in electricity). Use our cost calculator to model your own volumes.
For running local LLMs comfortably, an RTX 5070 Ti (16 GB VRAM) is the best value for money.
Amazon Check RTX 5070 Ti price →As an Amazon Associate, BestLLMfor earns from qualifying purchases, at no extra cost to you. It does not influence our independent rankings.