North Small Translate 1.0 vs NLLB-200: 50 Languages Against 200

The newer model covers a quarter of the languages. North Small Translate 1.0, which Cohere documented in its release notes on September 9, 2026, translates between English and 49 other languages. NLLB-200, the No Language Left Behind family Meta released in July 2022, covers 200. If language count were the whole comparison this would be a short article, but it is not: the two models are different machines built for different jobs, and the four-year gap between them shows up in ways that have nothing to do with how many languages appear on a list. What follows is what each one is actually good for, what they cost to run, and where the licensing sends you.

The coverage number, honestly

Two hundred versus fifty is real and it is the single most cited reason to keep NLLB-200 in a stack. Meta trained it on more than 18 billion sentence pairs with an explicit emphasis on low-resource languages, and it translates directly between any pair rather than pivoting through English — a design choice that matters enormously for, say, Bengali to Tamil, where an English-pivot system loses quality twice.

What the number hides is that the overlap between the two lists is the commercially useful part. North Small Translate 1.0's fifty languages include German, French, Spanish, Portuguese, Italian, Dutch, Polish, Czech, Japanese, Korean, Simplified and Traditional Chinese, Arabic, Hebrew, Hindi, Bengali, Tamil, Telugu, Thai, Turkish, Vietnamese, Indonesian, Malay, Russian and Ukrainian — the languages that account for the overwhelming majority of enterprise localization volume. NLLB-200 covers all of those too, plus roughly 150 more that North Small Translate 1.0 does not touch at all. So the question is not which list is longer. It is whether your traffic includes languages that only one of the two supports.

Two different machines

The architecture gap is the part that determines what you can build. NLLB-200's largest released checkpoint is a sparsely gated mixture-of-experts model of roughly 54.5 billion parameters, and its dense siblings run at 3.3B, 1.3B, and down to a distilled 600M. All of them are encoder-decoder sequence-to-sequence models in the M2M-100 lineage: you hand the tokenizer a source segment, it returns a translated segment. There is no chat template, no system prompt, no multi-turn structure, and the released checkpoints are built around segment-level translation: Meta's own model card states the model was trained with input lengths not exceeding 512 tokens, and warns that longer sequences degrade. The tokenizer's combined budget for source and target is 1,024 tokens. NLLB-200's card also describes it as a research model not released for production deployment, and notes it is not intended for document translation — the distinction that separates it most sharply from what Cohere built.

North Small Translate 1.0 is the opposite shape. It is a decoder-only sparse MoE with 218 billion total parameters and 25 billion active per token, 128 experts with eight active plus shared experts, and attention that alternates sliding-window layers (window 4,096, RoPE) with global attention layers that carry no positional embeddings. It runs behind a chat template with a default system instruction, and its window is 16K tokens of input and 16K tokens of output. That output budget is the tell: this is a model built to be handed a document and asked for a document back, not a model built to be handed a sentence.

• Total parameters — North Small Translate 1.0: 218B MoE (25B active) vs NLLB-200: 54.5B MoE flagship, 600M–3.3B dense

• Architecture — decoder-only causal MoE with chat template vs encoder-decoder seq2seq with no chat template

• Context — 16K in and 16K out vs segment-level, tokenizer budget 1,024 tokens combined

• Languages — 50 (English + 49) vs 200

• Direct pairs — English-centric tiering vs any-to-any without an English pivot

• License — CC BY-NC 4.0 vs CC BY-NC 4.0

• Minimum footprint — 1×B200 at W4A16, 2×B200 at FP8 vs ~37GB VRAM at 4-bit for the 54B, 600M runs anywhere

• Announced — September 9, 2026 (weights gated on Hugging Face since August 14) vs July 2022

What NLLB-200 still owns

Three things, and they are not sentimental. The first is the long tail: if you need Oromo, Tigrinya, or any of the roughly 150 languages outside North Small Translate 1.0's list, the decision is already made. The second is footprint at the small end — the distilled 600M checkpoint is 1.4 million downloads deep and runs on hardware that would not even load the Cohere weights, which makes it the only one of the two that fits an air-gapped or edge deployment without quantization tricks. The third is maturity: NLLB-200 has four years of tooling, quantization forks, and production war stories behind it, plus a peer-reviewed Nature paper from June 2024 describing how it was built. Cohere has published a release note.

The trade the other way is equally concrete. NLLB-200's model card describes it as a research model that was not released for production deployment, and its flagship 54B checkpoint is a heavyweight — roughly 350GB of storage, and on the order of 108–120GB of VRAM to serve uncompressed. Meta does not sell a hosted NLLB endpoint. Running it at scale is your problem, and it is a real one.

The licensing trap on both sides

This is the part that surprises people, because the two models carry the same license and it is not the one most teams assume. Both NLLB-200 and North Small Translate 1.0 are released under Creative Commons Attribution-NonCommercial 4.0. Neither is usable in a commercial product as published. NLLB-200's non-commercial restriction has been contentious enough that a community project called Open-NLLB exists specifically to try to make a commercially usable derivative, which runs straight into the license it is derived from. Cohere's route is cleaner: buy a commercial license and deploy through Model Vault, with the FP8 weights free on Hugging Face for non-commercial work only.

There is one asymmetry worth flagging. Cohere runs North Small Translate 1.0 on the free tier of its Chat V2 API, free for trial and production keys until rate limits are reached — so you can run a real evaluation at zero cost and only sign something if you ship. NLLB-200 gives you weights and nothing else.

Deploying either one

North Small Translate 1.0 ships in three checkpoints — BF16, FP8, and NVFP4 W4A16 — with published minimum hardware for each: four B200s or eight H100s for BF16, two B200s or four H100s for FP8, and a single B200 or two H100s for W4A16. Serving goes through vLLM with cohere_melody>=0.9.0, and Cohere recommends greedy decoding, matching its production configuration. Output is wrapped in <|START_TEXT|> and <|END_TEXT|> markers that are not registered as special tokens, so a naive skip_special_tokens=True will leave them in your output.

NLLB-200 deploys through the standard transformers or fairseq paths with a much wider tolerance for small hardware, since the 600M and 1.3B distilled checkpoints are the ones most people actually run. The 54B MoE is the one that requires planning.

The routing question this matchup actually represents

Neither North Small Translate 1.0 nor NLLB-200 is in OrcaRouter's catalogue today, so this is not a piece about picking one off our shelf. It is worth naming the shape of the problem anyway, because "one model for the tier-one languages, another for the long tail" is a routing policy — and a routing policy is the kind of thing that should live in configuration rather than in application code. OrcaRouter's routing DSL expresses exactly that as YAML plus CEL rules, so a language-pair rule can send European pairs to one model and fall back to another when a pair is not supported, and failover chains mean an upstream that starts returning errors does not become your outage. That is one key and one integration across a catalogue of more than 200 models at provider list price with 0% markup, rather than a second vendor contract for every translation model you decide to try.

Which one you should reach for

If your content strategy touches languages outside Cohere's fifty, keep NLLB-200 and treat that as settled — no amount of architectural modernity compensates for a missing language. If you are doing enterprise localization into the languages Cohere listed, work within a document rather than a sentence, and are willing to buy a commercial license, North Small Translate 1.0 is the more capable model on every dimension the documentation discloses, with the important caveat that its only quality evidence is one unreproduced vendor figure. If you are prototyping and have no budget, the Cohere free tier makes that evaluation nearly free, which is a better starting position than NLLB-200's, where the price of finding out is a GPU rental.

The useful way to hold both facts at once: NLLB-200 is the model that reaches the languages nobody else does, and it was built in 2022 for sentence-level research translation. North Small Translate 1.0 is the model that does fewer languages better, and it was built in 2026 for documents. Choosing between them is not really a quality judgment. It is a question about which languages you actually ship.