Tencent Hy-MT2-7B vs Tencent Hy-MT2-1.8B: The Server-Side Quality Pick or a 440MB Translator on Every Phone

Tencent Hy-MT2-7B and Tencent Hy-MT2-1.8B are the two bookends of the same family, open-sourced together on May 21, 2026, and both became callable through hosted APIs this week — the 7B around August 19 and the 1.8B a day later, each served by Tencent Cloud. They are not competitors in the usual sense, because they barely overlap in what they are for. The 7B is the quality pick: a dense model that runs on a single GPU or an API at $0.074 per million input and $0.295 per million output tokens. The 1.8B is the on-device pick: a model quantized down to 440 megabytes that runs on a phone, fully offline, at $0.044 / $0.177 per million on the API tier. If you are choosing between them, the answer is mostly determined by where the translation has to happen — and only secondarily by benchmarks, which are much closer than the fourfold parameter gap implies.

The one-line answer

Pick the 7B when translation happens in a data center and you want the best quality per server dollar — on an API or on a single A100-class GPU. Pick the 1.8B when translation has to happen on a device, offline, or at the lowest possible cost per token. If both live in the same cloud and budget is not the deciding factor, the 7B wins on quality. If the content is confidential, regulated, or needs to work without a network, the 1.8B is the only one of the two that can.

Specs, side by side

• Parameters — 7B dense vs 1.8B dense.

• Quantized footprint — FP8 and GGUF down to a few GB vs 440MB via AngelSlim 1.25-bit quantization.

• FLORES-200 (XX⇔XX) — 86.89 vs 79.77 XCOMET-XXL, roughly 97.9% vs 89.9% of Gemini 3.1 Pro (Think), vendor-reported.

• WMT25 — 7B: 63.86/71.21/82.24; 1.8B beats Microsoft and Doubao commercial APIs on all three metrics.

• DomainMTBench (GEMBA) — 92.79 vs 91.08, vendor-reported.

• API price — $0.074 / $0.295 vs $0.044 / $0.177 per 1M tokens (input/output).

• Context — both 8,192 tokens.

• Deployment — one A100 / RTX 4090 or cloud API vs Apple, Qualcomm, and MediaTek on-device chips, offline.

The quality gap is real but narrower than the size gap

All of the following is Tencent's own evaluation, unreproduced. On FLORES-200, the 7B sits an eight-point XCOMET ahead of the 1.8B (86.89 to 79.77), which is the gap the "balanced" versus "on-device" positioning is built on. But on DomainMTBench — the domain-translation benchmark covering finance, politics, and education — the 1.8B lands within 1.7 GEMBA points of the 7B (91.08 to 92.79). And the 1.8B's standing against the wider world is the detail that keeps surprising people: Tencent reports it beats Microsoft's and Doubao's commercial translation APIs across all three WMT25 metrics, and outperforms Tower-Plus-72B, a model forty times its size. So on ordinary domain text, the small model is far closer to its big sibling than four parameters would suggest. The 7B's real edge shows on the general-translation tail and on hard instructions, where its FLORES lead and its higher IFMTBench complex-task scores put clear water between the two.

The deployment fork

The 7B needs infrastructure — either a cloud API or a single A100-class GPU, with the usual ops attached. The 1.8B, at 440MB with AngelSlim's 1.25-bit quantization, runs on the same chips as a typical phone app, is 1.5× faster than the previous Hy-MT1.5 generation, and does not need a network at all. That turns privacy from a feature into a default: for contracts, medical records, or anything regulated, the data never leaves the device, which is a property no per-token price can buy on the server side. The trade is visible on FLORES-style general translation, where the 7B's extra capacity shows — and on any instruction complex enough to push the 1.8B's 83.14 IFMTBench total against the 7B's higher ceiling.

The cost math

On the API tier both models are cheap; the 1.8B is cheaper. At $0.044 / $0.177 versus $0.074 / $0.295 per million, the small model runs about 40% below the 7B on both sides of the bill — at a steady one million output tokens a day, roughly $70 per day versus $118. On-device, the 1.8B costs zero per token, which is the number that dominates any server comparison at volume. The 7B's counterargument is not price but headroom: if your corpus skews technical, legal, or instruction-heavy, the 7B's quality margin is exactly the kind of thing that shows up as fewer retranslations — and retranslations are the real unit cost in professional MT.

Routing the two sizes

Neither model is on OrcaRouter's catalogue as of this writing, so the honest framing is that both are served through Tencent's own cloud and several third-party platforms. The matchup still teaches the routing lesson this blog is built around: when two models overlap on capability but differ on price and latency, the rational deployment is not "pick one" but "route by workload" — the high-volume, low-difficulty slice to the cheap small model, the hard slice to the quality model, automatic failover so an outage in either never stalls the pipeline. That is precisely the pattern OrcaRouter's routing DSL composes across the 200+ models we do carry, with each provider's list price passed through at 0% markup. If Tencent's translation family joins the catalogue, it is the natural next tenant.

The verdict

If translation happens in your data center, take Tencent Hy-MT2-7B — on the API if you want zero operations, on a single GPU if you want to own it — and stop reading. If translation has to happen on a device, offline, or under a confidentiality requirement, take Tencent Hy-MT2-1.8B and the 440MB footprint wins by definition. The only genuinely hard case is a mid-volume cloud workload where the 7B's quality and the 1.8B's price are both defensible. There, do not decide on vendor benchmarks: the reported GEMBA gap is under two points, so run both on your own domain text and let your data pick.