Hy-SimulMT-30B-A3B vs North Small Translate 1.0: Two Big Models, One Licence You Can Ship

The shortest way to compare Hy-SimulMT-30B-A3B and North Small Translate 1.0 is to look at what happens after you decide you like them. Tencent's model is Apache-2.0: download it, fine-tune it, put it in a product, charge money, no conversation required. Cohere's is CC BY-NC 4.0: free for non-commercial use, gated behind a contact form, with commercial deployment routed through a paid licence from Cohere or its partner RWS. Both are large sparse mixture-of-experts models that most teams cannot casually run. Only one of them can be shipped on a Tuesday afternoon without a procurement call, and that asymmetry decides more deployments than any benchmark row will.

They also do not do the same thing, which makes the licence comparison sharper rather than softer. North Small Translate 1.0 was announced by Cohere on 10 September 2026 as a document-scale text translator: 218 billion parameters total, 25 billion active, a 16,000-token input window and — unusually — a 16,000-token output window, across 50 languages. Hy-SimulMT-30B-A3B, published as weights by Tencent on 22 October 2026, is a speech model: it listens to Chinese or English audio in thirty-second windows and produces the other language while the speaker is still talking. One translates a contract; the other translates a conversation. The overlap is smaller than the shared "translation model" label suggests.

Two different definitions of "small"

Cohere's name is relative to its own flagship. Against a frontier general model, 218 billion parameters with 25 billion active is small — it is built to be cheap per token and fast, deliberately non-reasoning, and Cohere reports about 112 output tokens per second. Against anything you would run on a workstation, it is enormous. Cohere's suggested hardware is one B200 or two H100s at W4A4; the model card's per-precision minimums are higher, with BF16 listed at four B200s or eight H100s. The FP8 weight set alone is on the order of 218 GB.

Tencent's "30B-A3B" means 30.4 billion total parameters and roughly 3 billion active per token, which is small in the only sense that matters for serving cost: each token activates about a tenth of the network. The BF16 weights are still around 61 GB before KV cache, so an 80 GB accelerator remains the practical floor, and the published FP8 build is what brings a smaller card into range. Neither model is small in the way a 440 MB on-device translator is small. The difference between them is one order of magnitude in total parameters and a factor of eight in active ones, which changes the hardware bill but not the category.

What each one does with its context

North Small Translate 1.0's headline feature is output length. Sixteen thousand tokens in and sixteen thousand out means a whole report, manual or contract can go in and come out translated as a unit, with terminology and tone consistent across the document because the model can see all of it at once. Cohere's long-context evaluation claims 48.9 against 21.3 for Google Translate and 19.4 for Gemma 4 31B on that kind of task. That is the capability that justifies the parameter count: not better sentences, but a better document.

Hy-SimulMT-30B-A3B has almost the opposite relationship with context. Its audio encoder accepts a maximum of 480,000 samples at 16 kHz — thirty seconds — per window, and its generation configuration defaults to sixty-four new tokens per step. The model is engineered to commit early and keep pace, not to hold a document in view. Its maximum position embedding count is 262,144, but the card publishes no usable context length, and for streaming audio the number is close to irrelevant. If your unit of work is a 40-page manual, this is the wrong tool by construction. If your unit of work is a sentence that is still being spoken, it is the only one of the two that applies.

The licence is the sharper difference

Apache-2.0 on Tencent's weights is about as permissive as open weights get: commercial use, modification, redistribution and derivative models are all permitted, with the standard carve-out that bundled third-party components keep their own terms. The licence file is explicit that Hy-SimulMT refers to the inference code, parameters and weights Tencent published, and that it imposes no restrictions beyond those of the components it ships. There is no acceptable-use addendum in the repository and no gate on the download.

Cohere's CC BY-NC 4.0 is a different instrument. Non-commercial use is free; commercial use requires a licence, and the weights are gated. Commentators have pointed out that "non-commercial" can be read broadly enough to cover monetised output — an ad-supported dubbed video, for instance — which makes the boundary a legal question rather than an engineering one. Cohere sells commercial access through its own sales channel and through RWS Language Weaver's Model Vault, and no public price is listed.

Put plainly: if you are building a product, Tencent's model is one you can evaluate and deploy on the same day, and Cohere's is one you can evaluate and then negotiate over. That does not make Cohere's model worse — a company with an existing RWS relationship may find the commercial path straightforward — but it does change the order of operations, and it means the two models are rarely shortlisted by the same team for the same reason.

The benchmarks cannot be set against each other

It is tempting to put Cohere's WMT26 score of 83.60 next to Tencent's 84.7 on RealSI-Sentence and declare a winner. Resist it. The two numbers come from different tasks, different metrics and different judges. Cohere's figure is from its own evaluation of the WMT26 all-languages set, scored by GPT-5.6-Sol as an automatic judge, with 84.36 for an agentic multi-pass workflow; the official WMT26 human evaluation had not been published when the model launched. Tencent's 84.7 is a speech-translation score on RealSI-Sentence, one column of a vendor-run table, alongside 90.3 FLEURS, 75.5 CoVoST2 and 81.2 ACL6060. Neither company has had its numbers reproduced by an independent lab.

What can be said is directional. Cohere published a comparative table with named competitors — DeepL NextGen at 81.37, Qwen 3.5 397B A17B at 81.56, Google Translate at 68.20 — all scored by the same judge, which makes the ranking internally coherent even if the judge is a language model. Tencent published a table of nine systems on four speech benchmarks, with its own model bolded in every column and a careful sentence noting it did not win CoVoST2. Both are marketing documents. The Cohere one is more useful for choosing between text translators; the Tencent one is the only evidence that exists for this speech model.

Deployment and cost, point by point

• Total parameters — Hy-SimulMT-30B-A3B: 30.4B, sparse MoE, roughly 3B active; North Small Translate 1.0: 218B total, 25B active.

• Weight footprint — Hy-SimulMT: ~61 GB in BF16, FP8 and GGUF builds published; North: on the order of 218 GB in FP8, with BF16 and NVFP4 variants.

• Hardware — Hy-SimulMT: one 80 GB accelerator realistically; North: one B200 or two H100s at W4A4 per Cohere's suggested configuration, higher for BF16.

• Context — Hy-SimulMT: thirty-second audio windows, no usable text context published; North: 16K in, 16K out.

• Languages — Hy-SimulMT: Chinese and English documented for speech, 36 tagged in metadata; North: 50.

• Licence — Apache-2.0, ungated vs CC BY-NC 4.0, gated, commercial licence required.

• Hosted API — neither is on OrcaRouter's catalogue and we host neither. Tencent publishes no endpoint or price for Hy-SimulMT; Cohere offers commercial access through Model Vault without a public rate card.

Where a router earns its keep in this kind of build is downstream of the translation itself. A document pipeline that uses North, or a meeting pipeline built on Hy-SimulMT, still needs a reviewer: a second model that checks terminology, flags omissions and compares the translation against the source. Running that check as a separate call to a different vendor is exactly the kind of multi-model workflow OrcaRouter is built for — one API for 200+ models, 0% markup on provider list prices, failover if the reviewing provider has a bad minute, and model fusion when you want a panel rather than a single critic. The translator is one node; the quality control around it is where the routing decisions live.

How to choose

If the job is translating documents, the decision starts and ends with the licence. A team that can buy a commercial licence and needs 16K of output should look hard at North Small Translate 1.0 — the output window is a real capability that most translation models do not have. A team that needs to ship without a negotiation, or that wants to fine-tune and redistribute, has exactly one option here, and it happens to be the one that costs a tenth of the parameters.

If the job is live speech, the comparison collapses. North Small Translate 1.0 does not accept audio, has no streaming mode, and a three-second lag claim is meaningless for a model that returns a document. Hy-SimulMT-30B-A3B is the only one of the two that can sit in a meeting, and its Apache-2.0 licence means you can find out whether it works without asking anyone's permission. That is a narrow verdict, but it is an honest one: these models were built for different rooms, and the licence just makes the choice starker.