North Small Translate 1.0 vs Hy-MT2-7B: One GPU Against Two B200s

The closest matchup in open translation right now may also be the least obvious one. Hy-MT2-7B is the middle child of Tencent Hunyuan's translation family, open-sourced under Apache 2.0 on May 21, 2026, and it runs on a single RTX 4090 or A100 in about 16GB of VRAM. North Small Translate 1.0 is Cohere's 218-billion-parameter sparse mixture-of-experts translator, documented in the company's release notes on September 9, 2026, with a minimum deployment of two B200s at FP8 or one B200 in its NVFP4 W4A16 form. Both are translation specialists. Both are current. One of them you can serve from a workstation, and that single fact reframes the entire comparison — because the benchmark where they would most want to be judged on equal terms does not exist.

Start with the envelope, not the scores

Deployment cost is the dimension that decides most real integrations, and here the two models are not close. Hy-MT2-7B is a dense HunYuanDenseV1 model of roughly eight billion parameters with published FP8 and GGUF builds, plus community MLX 8-bit versions for Apple silicon. Tencent's deployment notes list transformers 5.6.0 or later, vLLM, SGLang and llama.cpp as supported runtimes, and the inference guidance caps generation at 8,192 tokens even though the configuration advertises up to 262,144 positions. A single modern consumer or workstation GPU covers it.

North Small Translate 1.0 is a different class of object. Its 218B total parameters with 25B active are split across 128 experts with eight active per token plus shared experts, and Cohere publishes minimum hardware for each of its three checkpoints: four B200s or eight H100s for BF16, two B200s or four H100s for FP8, one B200 or two H100s for the NVFP4 W4A16 build. Serving runs through vLLM with cohere_melody>=0.9.0, and Cohere recommends greedy decoding to match production. Output carries <|START_TEXT|> and <|END_TEXT|> markers that are not registered as special tokens.

• Shape — North Small Translate 1.0: 218B total / 25B active sparse MoE, 128 experts vs Hy-MT2-7B: dense, roughly 8B

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

• Minimum hardware — 1×B200 at W4A16, 2×B200 or 4×H100 at FP8 vs a single RTX 4090-class GPU in ~16GB

• Published formats — BF16, FP8, NVFP4 W4A16 vs base, FP8, GGUF, plus community MLX 8-bit

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

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

• Reported quality — WMT26 83.60, metric unnamed, vendor-reported vs named vendor tables on FLORES-200, WMT25 GEMBA, XCOMET-XXL and IFMTBench

The benchmark problem

Here is the honest core of this matchup: we cannot rank these two models against each other, and anyone who tells you otherwise is inventing a number. Tencent published four named evaluations. On FLORES-200 English-to-X translation the 7B posts 93.52, behind Gemini 3.1 Pro's 94.42 and GPT-5.5's 94.16 but close enough to matter. On the WMT25-adjacent GEMBA metric it scores 82.24 against the 30B-A3B's 84.34 and GPT-5.5's 83.41. On XCOMET-XXL it leads everything in Tencent's comparison table at 63.86 — ahead of Gemini 3.1 Pro, GPT-5.5, DeepSeek-V4-Pro and Kimi K2.6. On IFMTBench's instruction-following score it lands at 83.14%. All of these are Tencent's own numbers, none have been independently reproduced, and they are still far more than Cohere has offered.

Cohere published a single figure: a WMT26 score of 83.60, rising to 84.36 under an agentic multi-pass workflow. No metric is named on the model card or in the release note, and no WMT26 results page has been published against this model. That asymmetry is not proof that Cohere's model is weaker or stronger. It means the comparison a reader most wants — same test, same metric, both models — cannot be made from public evidence, and the four-point spread inside Cohere's own two figures (83.60 to 84.36) is already larger than most of the gaps in Tencent's table.

Languages and the long document

North Small Translate 1.0 covers fifty languages and locale variants, with Modern Standard Arabic, German, French, Japanese, Korean, Russian and Ukrainian designated tier-one, and it accepts and emits 16,000 tokens on both sides. Hy-MT2-7B covers thirty-three languages plus five minority languages and dialects including Tibetan, Uyghur and Cantonese. Neither list is a superset of the other — Tencent reaches Cantonese and Uyghur, Cohere reaches a broader set of European locales — so a multilingual deployment may find that it needs both on coverage grounds alone.

The output window is the quieter difference. Sixteen thousand tokens of output means a full procedure document in one pass, with terminology and register consistent across the whole thing because the model saw all of it. An 8K window does not, and the sentence-by-sentence workaround reintroduces exactly the cross-segment consistency problems that translation instruction-following benchmarks like IFMTBench were built to measure.

The license, again, decides more than the table

North Small Translate 1.0 is CC BY-NC 4.0. The weights are free for non-commercial work, and commercial deployment runs through Cohere's Model Vault under a purchased license with no published price. Hy-MT2-7B is Apache 2.0 and ungated — commercial use, fine-tuning and derivatives all permitted without a conversation. For a commercial product the order of operations writes itself: Hy-MT2-7B is deployable today and Cohere's model is evaluable today, and no benchmark row changes that ordering.

Cohere's offsetting advantage is the free tier. North Small Translate 1.0 runs on the free tier of Chat V2, free for both trial and production keys until rate limits are reached, so an evaluation costs nothing but time. Hy-MT2-7B has a hosted commercial API — Tencent priced the family's hosted tier at roughly $0.044 per million input tokens and $0.177 per million output tokens — and self-hosting on your own GPU is the genuinely free path.

What "multi-pass" actually implies

Cohere's decision to publish both an 83.60 and an 84.36 is the most informative detail in its release note, because it says the company believes a workflow beats a single call. That is the same bet Tencent made with a different mechanism: its flagship 30B-A3B wins on GEMBA and XCOMET while Gemini 3.1 Pro wins on FLORES-200, which means the best system is not one model but a panel. OrcaRouter's model fusion runs several models as a panel and reconciles their answers within a single call, which is the platform-level version of the multi-pass idea both vendors are chasing — and it composes with the routing side, where one key covers a catalogue of more than 200 models at provider list price with 0% markup, so a vendor price change lands on our side the same day rather than at the next contract renewal.

That framing also solves the evidence problem this article opened with. When two vendors publish non-overlapping benchmarks and neither has an independent evaluation, the way to choose is not to trust a table. It is to run both on your own documents and let the disagreement surface on real text — which is exactly what a panel and a failover chain are for.

Which one, and when

If you need a translation model running on hardware you already own, in a commercial product, without a licensing conversation, Hy-MT2-7B wins on the strength of the envelope alone — and its vendor-published results, unreproduced as they are, are at least named, comparable and close to the frontier. If you are translating long documents into Cohere's fifty languages, need 16K of consistent output per pass, and have either a two-B200 budget or a willingness to run the evaluation on Cohere's free tier before deciding, North Small Translate 1.0 is the more capable machine on every disclosed axis.

What neither model does is remove the need to test. Cohere has given you one unnamed number and a free way to check it. Tencent has given you a table and a GPU-sized download. The 83.60 is a promise, not a result — and the only place that promise gets settled is on your own corpus.