North Small Translate 1.0 vs Hy-MT2: Two MoE Translators, One Evidence Gap

Both of these families were built on the same bet — that a sparse mixture-of-experts network post-trained for translation beats a general-purpose model asked to translate — and they arrived at it from opposite directions. Hy-MT2, Tencent Hunyuan's three-model translation family headlined by the Hy-MT2-30B-A3B MoE, was open-sourced on May 21, 2026 under Apache 2.0 with a technical report, an open-sourced benchmark and a full comparison table. North Small Translate 1.0, Cohere's 218-billion-parameter, 25-billion-active MoE translator, was documented in a release note dated September 9, 2026, and its quality evidence amounts to a single number with no metric attached to it. The architectures rhyme. The documentation does not, and that difference is the most useful thing to understand before choosing either.

One family, three sizes against one large model

Hy-MT2 is not a single model but a range: a 1.8B dense model for on-device work, a 7B dense model for a single GPU, and the 30B-A3B MoE as the flagship with three billion parameters active per token. All three are Apache 2.0, ungated, and released together with FP8, GGUF, 2-bit and 1.25-bit quantizations plus the IFMTBench instruction-following benchmark and a companion paper. Tencent reports that the family translates among 33 languages plus five minority languages and dialects.

North Small Translate 1.0 is a single point in the size space, and a large one: 218B total parameters, 25B active, 128 experts with eight activated per token plus shared experts, and attention alternating in a 3:1 ratio between sliding-window layers (window 4,096, RoPE) and global layers carrying no positional embeddings. Its window is 16K in and 16K out across English plus 49 other languages, with Modern Standard Arabic, German, French, Japanese, Korean, Russian and Ukrainian designated tier-one. Notably, the shape is not new — Cohere's Command A+ from May 2026 used the same 218B/25B configuration — which makes this a translation-specialized post-train of an existing core rather than a new architecture.

• Parameters — North Small Translate 1.0: 218B total / 25B active vs Hy-MT2-30B-A3B: 30B total / 3B active, with 1.8B and 7B dense siblings

• Context — 16K in and 16K out vs 8K working window, config advertises up to 262,144 positions

• Languages — 50 (English + 49) vs 33 plus 5 minority languages and dialects

• License — CC BY-NC 4.0, commercial via Model Vault vs Apache 2.0, ungated

• Published evidence — one unnamed WMT26 figure, vendor-reported vs a technical report, an open benchmark, and named results on FLORES-200, WMT25 GEMBA and XCOMET-XXL

• Serving — vLLM with cohere_melody>=0.9.0, greedy decoding recommended vs transformers, vLLM, SGLang and llama.cpp

• Announced — September 9, 2026 (weights gated on Hugging Face since August 14) vs May 21, 2026

The evidence gap is the real story

Tencent's release came with receipts. There is a paper on arXiv, a benchmark the company wrote and open-sourced specifically to measure translation instruction-following, and a results table naming its competitors. FLORES-200 English-to-X puts the 30B-A3B at 93.85 against Gemini 3.1 Pro at 94.42 and GPT-5.5 at 94.16. The WMT25-era GEMBA metric puts it at 84.34 — the highest score in Tencent's own comparison, ahead of GPT-5.5 at 83.41 and 83.29 in its two modes, Gemini 3.1 Pro at 82.23, DeepSeek-V4-Pro at 81.99 and Kimi K2.6 at 81.68. XCOMET-XXL places it at 62.89 behind its own 7B sibling. On IFMTBench it reaches 85.7% overall instruction-following, with a sub-score of 91.94% that sits within a hundredth of a point of Gemini 3.1 Pro and a separate sub-score, 85.55%, where it leads the Gemini 3.1 Pro model outright.

Every one of those is Tencent's own measurement and none has been independently reproduced. But they are named, they are comparable, and they are falsifiable — a reader knows exactly which test to run to argue.

Cohere's release note offers one figure: WMT26 83.60, rising to 84.36 under what it calls an agentic multi-pass translation workflow. No metric is named anywhere on the model card or in the docs. No WMT26 results page has been published against this model. The training corpus, token count and data pipeline are undisclosed, where Command A Translate's 2025 technical report had described a difficulty-filtering technique in detail. None of that is evidence of a worse model. It is evidence of less evidence, which is a different thing and matters just as much when you are deciding what to put in production.

The shared yardstick neither side has actually published

There is one genuine point of contact between the two, and it is worth a paragraph because it is the only place a fair comparison could happen. Tencent is a WMT26 partner, sponsoring a video subtitle translation task with Hunyuan-funded awards. Cohere's single published number is a WMT26 figure. So both organizations are inside the same 2026 evaluation cycle, and the one metric where a head-to-head could be settled is precisely the one where only one of them has published anything — and published it without naming the metric.

That is the gap to watch. If Cohere attaches a metric name to the 83.60, or if the WMT26 results pages include a North Small Translate system, the comparison becomes real. Until then, stacking Tencent's FLORES-200 and GEMBA numbers against Cohere's unlabeled WMT26 figure would be a fabrication dressed up as analysis.

License and deployment

Hy-MT2 is Apache 2.0 with no gating: commercial use, fine-tuning, derivatives and redistribution, all without a conversation. North Small Translate 1.0 is CC BY-NC 4.0, free for non-commercial use, with commercial deployment routed through Cohere's Model Vault at a price that is not published. For anything shipping to customers, that difference resolves the decision before the benchmarks are read.

The hardware story follows the same pattern in reverse. Hy-MT2 spans a 440MB quantized build that runs on a handset up to the 30B-A3B, whose three billion active parameters keep per-token compute modest for its quality tier; runtimes cover transformers, vLLM, SGLang and llama.cpp. North Small Translate 1.0 publishes minimum hardware per checkpoint — four B200s or eight H100s for BF16, two B200s or four H100s for FP8, one B200 or two H100s for NVFP4 W4A16 — and recommends greedy decoding to match production. Cohere's offsetting advantage is that the model runs on the free tier of its Chat V2 API, free for trial and production keys until rate limits are reached, so evaluating it costs nothing.

Should you switch, and to what

The switch question is genuinely asymmetric here. Moving from Hy-MT2 to North Small Translate 1.0 is not a performance upgrade you can currently justify — it is a bet on a model whose only public claim is one number, traded against a family with a published report and a deployable license. What it is worth doing is evaluating: the Cohere model is free to call, covers fifty languages including a broader European set, and gives you 16K of output per pass, which Hy-MT2's 8K working window does not.

That evaluation is the case where a routing layer earns its keep, because the honest answer for most multilingual stacks is that both models stay. OrcaRouter puts a catalogue of more than 200 models behind one key, so trying a second translation model is a configuration change and not a second vendor contract or a code change — you point the same OpenAI-compatible client at a different model id and compare on your own text. Provider list price is passed through at 0% markup, so a hosted price change is live on our side the same day with no spread layered on top, and failover chains keep production on the model that already works while the new one is being graded. Neither North Small Translate 1.0 nor Hy-MT2 is in our catalogue today, so this is not a recommendation to buy either from us — it is the argument for not hard-coding the decision before you have run the test.

The verdict

Tencent's Hy-MT2 is the safer choice and it is not close on the evidence: Apache 2.0, three sizes, a paper, an open benchmark, four named evaluation suites and a WMT26 partnership. Cohere's North Small Translate 1.0 is the more interesting one — 218 billion parameters of translation-specific MoE behind a 16K output window and fifty languages, a free evaluation tier, and an 83.60 that nobody outside Cohere has checked.

If you need a decision this quarter, take the family with the receipts. If you have a corpus and a week, put some of it through the free tier and find out whether Cohere's unnamed 83.60 means anything on your documents — because that is a question no vendor's table will answer for you, and the one number Cohere chose to publish is precisely the one it declined to name.