North Small Translate 1.0 vs Hy-MT2-1.8B: 218GB of Model Against 440MB

One of these fits on a phone. Hy-MT2-1.8B, Tencent Hunyuan's on-device translation model open-sourced under Apache 2.0 on May 21, 2026, shrinks to 440 megabytes under AngelSlim 1.25-bit quantization and runs on Apple, Qualcomm and MediaTek phone silicon. North Small Translate 1.0, the 218-billion-parameter mixture-of-experts model Cohere documented in its release notes on September 9, 2026, ships an FP8 weight set on the order of 218 gigabytes and asks for two B200s at minimum. That is roughly five hundred times the storage for one of them, and the interesting question is not which is better — it is what each one is actually for, and which license lets you ship it.

The size gap is not a quality gap

It is tempting to read 218B against 1.8B as a straight capability ranking. It is not, for two reasons. The first is that North Small Translate 1.0 is a sparse mixture-of-experts model with only 25 billion parameters active per token, so the compute per token sits in a different league from the parameter count. The second is that the two models were optimized against different objectives: Tencent's 1.8B was built to be small enough to run offline on a handset, and Cohere's model was built to hold a whole document in context and translate it as a unit.

That difference is legible in the context windows. Hy-MT2-1.8B's configuration advertises up to 262,144 positions, but Tencent's own inference guidance caps generation at 8,192 tokens — the practical working window is 8K. North Small Translate 1.0 documents 16K in and 16K out, which is double the input window and, more importantly, a full 16K of output. If your unit of work is a service manual, that output budget is the feature. If your unit of work is a chat message, it is irrelevant.

• Total parameters — North Small Translate 1.0: 218B MoE, 25B active vs Hy-MT2-1.8B: 1.8B dense

• Context — 16K in and 16K out vs 8K working window (config advertises up to 262,144 positions; Tencent's guidance says 8,192)

• 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

• Quantized footprint — FP8 ~218GB, NVFP4 W4A16 ~109GB vs 440MB at 1.25-bit, FP8 and GGUF also published

• Minimum hardware — 2×B200 or 4×H100 at FP8 vs a handset

What Tencent actually shipped in the 1.8B

The 1.8B is the smallest member of a three-model family — 1.8B, 7B, and a 30B-A3B MoE flagship — all released together with a companion benchmark called IFMTBench that measures translation instruction-following rather than raw translation quality. Tencent shipped FP8, GGUF, 2-bit and 1.25-bit quantizations alongside the base weights, and the 1.25-bit version is the one that produces the 440MB figure and a claimed 1.5x inference speedup over the previous Hy-MT1.5 generation. Serving is supported through transformers 5.6.0 and later, vLLM, SGLang and llama.cpp, with the GGUF path depending on an STQ kernel landed in llama.cpp. Recommended sampling is temperature 0.7, top_p 0.6, top_k 20, repetition penalty 1.05, and there is no default system prompt — the inverse of Cohere's approach.

It is also, unlike North Small Translate 1.0, a model with visible adoption. The Hugging Face repository has been downloaded more than 23,000 times and carries around 1,200 likes. Cohere's main repository showed 26 downloads and zero likes when this was written.

The license is the sharper difference

This is the part that should decide more deployments than the benchmark table does. Hy-MT2-1.8B is Apache 2.0 with no gating — commercial use, fine-tuning, derivative works, redistribution, all permitted. North Small Translate 1.0 is CC BY-NC 4.0: the weights are free for non-commercial use, and commercial deployment requires a purchased license through Cohere's Model Vault, for which no price is published.

Put plainly: if you are building a product, Tencent's model is the one you can ship today without a conversation, and Cohere's is the one you can only evaluate. That asymmetry does not make the Cohere model worse, but it changes the order of operations — you evaluate Cohere's, and you can deploy Tencent's.

Quality evidence is asymmetric, and both sides are vendor-reported

Neither model has an independent evaluation behind it, so treat every number here as a claim by its maker. The difference is how much the makers have put on the table. Tencent published a full comparison table and a technical report: on FLORES-200 English-to-X translation, Hy-MT2-1.8B posts 90.00 against Gemini 3.1 Pro's 94.42 and GPT-5.5's 94.16 — modestly behind the frontier models but within a few points, from a model that fits on a phone. On IFMTBench's overall instruction-following score the 1.8B lands at 69.36%, well behind its own 30B-A3B sibling at 85.7% and Gemini 3.1 Pro at 89.08%, which is the honest read on the tradeoff: the tiny model follows formatting and terminology instructions far less reliably than it translates.

Cohere published one figure — a WMT26 score of 83.60, rising to 84.36 with an agentic multi-pass workflow — with no metric name attached and no WMT26 results page to check it against. That is not a comparison we can make. A single unnamed number cannot be set against a table of named benchmarks, and pretending otherwise would be the most misleading thing this article could do.

Tencent's counterweight is that its numbers come from somewhere a reader can inspect. The Hy-MT2 repository publishes the family overview, the three model sizes and the runtimes each one supports alongside the benchmark table — which is what makes the FLORES-200 and IFMTBench figures checkable at all, and what makes Cohere's single unnamed number conspicuous by contrast.

What each one costs to call

Tencent's 1.8B has a hosted commercial API that launched in August 2026, priced at roughly $0.044 per million input tokens and $0.177 per million output tokens — cheap in absolute terms, and priced per token. Cohere's model runs on the free tier of Chat V2, free for trial and production keys until rate limits are reached, so the evaluation cost is zero but the production cost is a contract you have to negotiate. Self-hosting flips the arithmetic entirely: 440MB on a handset versus two B200s, which is a difference measured in orders of magnitude rather than percentages.

The reason that gap is worth stating in a piece about a routing platform is that the cheap tier and the expensive tier are not competitors — they are two positions in a cascade, and cascades are what a router is for. OrcaRouter passes provider list price through at 0% markup, so a vendor price cut on a hosted translation endpoint is live on our side the same day with no reseller spread to renegotiate, and failover chains let a smaller model absorb the bulk of the traffic while a heavier model takes the requests it fails. Try the cheap one first, escalate on the hard cases, and let the routing layer hold the policy rather than your application code.

Which one you should pick

If your translation happens on a device, offline, under a per-megabyte storage budget, or inside a commercial product with no appetite for a license negotiation, Hy-MT2-1.8B is the answer and North Small Translate 1.0 is not in the running. If your unit of work is a long document in one of the fifty languages Cohere supports, if you need 16K of output in a single pass, and if a commercial license is something your organization is willing to buy, then Cohere's model is the more capable machine on every disclosed dimension — with the caveat that its quality rests on a single unreproduced figure.

And for most teams the honest answer is both, in that order: the 440MB model does the routine work at trivial cost, the 218B model handles the documents that matter, and the decision about which request goes where is a routing rule rather than a rewrite. The gap between these two models is not a gap to be closed. It is a spectrum to be used.