translate step and serves three very different jobs: document localization, subtitle translation against timed dialogue, and dubbing script generation where the output has to fit a timing budget and sound speakable.
The catalog spans purpose-built machine translation (DeepL, Amazon), frontier general-purpose LLMs (OpenAI, Claude, Gemini), efficient open-weight models (Llama, Qwen, Mistral), and Ollang’s own agentic localization workflow. They are not interchangeable — they differ in language coverage, context window, steerability, and which Order Types they serve.
Choosing by job
Language coverage across this catalog ranges from 12 to over 240. Check the family’s coverage before committing a Workflow — it is the most common reason one has to be reconfigured.
Ollang Agentic
Ollang’s own localization workflow, and the only provider in this catalog that is a pipeline rather than a single model. It selects and combines agents per segment based on content type, domain, and structural complexity, then reconciles their output.
Agentic Localization covers end-to-end localization of text, documents, subtitles, and dubbing scripts across 100+ languages. It is the default because localization quality — not translation accuracy alone — is what most Orders are judged on.
Agentic Turbo is the same shape of workflow tuned for throughput, aimed at high-volume pipelines that still need contextual quality rather than raw machine translation.
OpenAI
The widest model ladder in the catalog. The current generation is GPT-5.6, which ships as Luna, Terra, and Sol in ascending capability. All three carry a 1.05M-token context window and 128K max output — the tiers differ in capability, not in how much text they can hold. Ollang exposes the two ends of that range.
All OpenAI models serve Document, Subtitle, and AI Dubbing Orders.
GPT-5.6-Sol is OpenAI’s most capable model, for complex translation and localization where nuance, terminology accuracy, consistency, and reliable instruction-following all have to hold at once. It supersedes the earlier 5.x tiers on capability.
GPT-5.6-Luna is the efficiency tier of the same generation, built for everyday translation at volume with minimal latency — while carrying the same 1.05M context window as Sol. That combination is unusual: whole-document context on a throughput-oriented tier.
GPT-5.4, GPT-5.2, GPT-5, GPT-4.1, and o4-mini are earlier tiers, retained for Workflows already benchmarked against them. GPT-5.4 brought roughly 1M-token context and configurable reasoning effort; GPT-4.1 remains a sensible mid-tier workhorse; o4-mini is the compact reasoning option with an efficient tokenizer for non-English text.
Anthropic Claude
Consistently the strongest family for long-form coherence, register control, and terminology consistency across a document — the qualities that show up in human review rather than in automatic scores. The current lineup all carry a 1M-token context window with 128K max output and adaptive thinking. On Anthropic’s current tokenizer, 1M tokens is roughly 555,000 words.
All Claude models serve Document, Subtitle, and AI Dubbing Orders, and all support text and image input.
Claude 5 Fable is Anthropic’s highest-capability model. Its thinking is adaptive and always on, and its comparative latency is the slowest in the lineup. Reserve it for genuinely hard content — dense technical or legal source, structurally ambiguous material, or work where a second human pass would otherwise be needed.
Claude 5 Opus offers the same 1M context and 128K output with the most recent knowledge cutoff in the family, and is the sensible default within Claude for quality-critical localization.
Claude 5 Sonnet is positioned by Anthropic as the best combination of speed and intelligence. If a Workflow currently runs Claude 4 or 4.5 Sonnet, this is a straight upgrade.
The 4.x tiers remain available for existing benchmarks, each now matched or beaten by its Claude 5 equivalent.
Google Gemini
The most natively multimodal family in the catalog — text, images, audio, video, and PDF as first-class inputs — with million-token context windows across the current generation.
All Gemini models serve Document, Subtitle, and AI Dubbing Orders.
Gemini 3.7 Flash is now a stable release rather than a preview. It carries a 1M-token context window with 64K output, accepts text, images, audio, and video, and has a knowledge cutoff of March 2026 — the most recent of any model in this catalog. It is the default recommendation within Gemini for most localization work.
Gemini 3.1 Pro Preview is the reasoning flagship, with a 1M-token input window and a three-tier thinking control that trades latency against reasoning depth. Suited to very large documents and mixed-media source material processed in a single pass.
Gemini 2.5 Pro remains stable and available. Gemini 3 Pro Preview has been superseded in Google’s current model list by the 3.1 and 3.x Flash releases; it is retained here for Workflows already benchmarked against it.
DeepL
Purpose-built neural machine translation rather than a general-purpose LLM, across 240+ languages and variants — the widest language coverage in the catalog.
DeepL is more steerable than “NMT” suggests. Its API supports glossaries (up to five per request), translation memory with adjustable matching thresholds, style rules, formality control, XML and HTML tag handling, a context parameter for text that influences the translation without being translated, and custom instructions — up to ten per request, 300 characters each. A
model_type parameter selects between quality-optimized and latency-optimized behavior.
Document translation preserves formatting across docx, pptx, xlsx, pdf, html, txt, xliff, srt, idml, xml, json, dita, and mif.
DeepL Extended adds translation of images embedded inside documents — the difference between a translated PDF whose diagrams are still in the source language and one that is genuinely localized.
Amazon Translate
Neural machine translation across 75 languages and variants, with the enterprise controls that come with AWS.
Its case rests on four output controls rather than on model capability: Custom Terminology, which forces specific renderings of brand, product, and industry terms; Active Custom Translation, which adapts the engine using your own parallel data; Formality, which selects register in the languages that support it; and Brevity, which constrains output length — genuinely useful for subtitle work, where character limits bind and expansion into a longer target language breaks reading rates. Profanity masking is available where content standards require it.
For organizations with a mature translation memory and hard terminology requirements, that combination is often worth more than a higher-capability general model that has to be re-steered on every Order.
Cohere
Enterprise-oriented models built for secure deployment, low latency, and predictable behavior in regulated environments.
Command A carries a 256K-token context window and is built for tool use, agents, retrieval-augmented generation, and multilingual work. It covers the 23 languages of global business: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic, Chinese, Russian, Polish, Turkish, Vietnamese, Dutch, Czech, Indonesian, Ukrainian, Romanian, Greek, Hindi, Hebrew, and Persian.
Its strongest fit is cross-lingual RAG — pipelines where a translation must be grounded in and traceable to an internal source of truth. Its Arabic, Hebrew, Persian, and Central and Eastern European coverage is a genuine strength relative to models that treat those as long tail.
Cohere also publishes Command A Translate (
command-a-translate-08-2025), a dedicated machine translation model covering the same 23 languages. Ollang’s cohere/command-a provider runs the general Command A model.Mistral
Mistral Medium 3.1 delivers translation and content generation across 50+ languages with a 130K-token context window and multimodal text-plus-image input. Its position here is efficiency — an instruction-following LLM tuned for throughput, which makes it a capable subtitle-translation workhorse for European language pairs.
Qwen
Qwen3-235B-A22B is a mixture-of-experts model: 235B total parameters, 22B active per token, routing across 128 experts with 8 activated at a time.
It carries a 262,144-token native context window — native rather than extended, so long-document behavior is what the model was trained for — and the Qwen3 family is trained across 119 languages and dialects. Qwen reports substantial gains in long-tail knowledge coverage across multiple languages, which is the characteristic that matters for translation outside the top 20.
The clearest reason to select it is Chinese–English work and long-tail language coverage.
Llama
Llama 4 does not serve AI Dubbing Orders.
Related
Provider catalog overview
Pipeline steps, provider tags, and status definitions.
Memory and custom instructions
Steer any of these models with glossaries, guidelines, and translation memory.