Tencent Hy-MT2-7B Now Has a Hosted API: What a $0.295-per-Million-Output Translator Changes

Tencent Hy-MT2-7B, the open-source translation model Tencent Hunyuan released on May 21, 2026, became callable through a hosted API this week: around August 19, 2026, the dense 7B model went live on Tencent Cloud's hosted endpoint at $0.074 per million input tokens and $0.295 per million output tokens, with an 8,192-token context window. It did not arrive alone — Tencent Hy-MT2-1.8B and Tencent Hy-MT2-30B-A3B came up on the same backend within a day. The weights have been on Hugging Face and ModelScope for three months; what is new is that a team can now call the thing without renting a GPU, building llama.cpp from source, or shipping the 1.25-bit-on-device toolchain. For a translation workload, that is the difference between an experiment and a product decision.

What just shipped

The commercial endpoint is Tencent Cloud's own, exposed through the vendor's API and several third-party platforms. The three sizes each run in an 8K context with 4,096-token completions; all three accept structured output via a JSON schema in their response_format field, and none of them implement function calling. The practical spec, in one line per dimension:

• Tencent Hy-MT2-7B — $0.074 in / $0.295 out per 1M tokens, 8,192 context, dense ~7B, structured output supported, no tool calls.

• Tencent Hy-MT2-1.8B — $0.044 in / $0.177 out per 1M tokens, 8,192 context, dense 1.8B, no tool calls.

• Tencent Hy-MT2-30B-A3B — $0.074 in / $0.295 out per 1M tokens, 8,192 context, MoE 30B total / 3B active, structured output supported, no tool calls.

The headline price point is the 7B's output rate: under three-tenths of a cent per thousand output tokens, for a model Tencent claims holds its own against models an order of magnitude bigger. Translation is one of the few LLM workloads where output tokens are nearly the whole bill, so output price is the number that determines whether a pipeline is viable at volume.

The model behind the endpoint

Hy-MT2 is Tencent's "fast-thinking" family: rather than spending long chains of thought on every sentence, it is designed to react the way a professional translator does — deliberate where the source is ambiguous, quick where it is not. The 7B is the balanced middle of the family, a dense model licensed Apache-2.0, covering 33 languages plus five Chinese minority-language and dialect pairs (Tibetan, Kazakh, Mongolian, Uyghur, Cantonese among them). What separates it from a generic LLM doing translation is instruction-following: it will hold a glossary term across a document, apply a specified style, leave delimiters and JSON keys untouched, and take personalization instructions in plain language. Tencent shipped IFMTBench, an open benchmark for exactly that skill, alongside the weights, and the consumer face — the Tencent Hy Translate mini-program, with iOS and Android apps in the pipeline — is built on this family.

The numbers, with the asterisk attached

Everything below is Tencent's own evaluation, published on the model card and in the accompanying report; none of it has been independently reproduced at this writing. On FLORES-200's XX⇔XX general-translation track, the 7B scores 86.89 XCOMET-XXL, which Tencent puts at roughly 97.9% of Gemini 3.1 Pro (Think). In its "fast-thinking" mode it is claimed to outrun open-source general models including DeepSeek-V4-Pro, Kimi K2.6, Qwen3.5-397B-A17B, and Gemma4-26B-A4B. On WMT25 its scores improve across the board over the previous generation — 63.86/71.21/82.24 versus Hy-MT1.5-7B's 61.59/68.85/75.91, with the biggest gain on GEMBA — and on DomainMTBench (finance, politics, education) it posts 94.92 XCOMET / 92.79 GEMBA. Its instruction-following is where it most visibly separates from translation LMs of a year ago: 89.73 simple / 72.67 complex / 83.14 total on IFMTBench.

What a hosted API changes for a translation pipeline

Self-hosting the 7B is genuinely easy as far as these things go — it runs on a single A100 or RTX 4090, and quantized FP8 and GGUF builds exist. But "easy to self-host" is not "zero operations." The GGUF path currently depends on llama.cpp with Tencent's STQ kernel, the FP8 path needs the right stack, and someone has to babysit it. A hosted endpoint moves all of that off the table: no GPU, no kernel build, no capacity planning, and a per-token price that is flat regardless of utilization. For a team processing millions of translated sentences a day, that predictability matters more than the headline benchmark — and it is the reason this week matters more than the May launch did.

The routing question nobody is asking yet

Because the model is three days old on the API side, the realistic way to adopt it is the way you adopt any brand-new provider: put it behind a routing rule with automatic failover so an untested endpoint can never take down a production path, and let the volume decide whether it earns a permanent slot. That is precisely the pattern OrcaRouter's routing DSL composes across the 200+ models we do carry — and it is worth being precise here: Tencent Hy-MT2-7B is not on OrcaRouter's catalogue as of this writing, so this is not a "call it through us" story. What is on-topic is the pricing model. OrcaRouter passes through each provider's list price with 0% markup, so when a vendor cuts the rate, the cut is live on the platform the same day. For a high-volume translation workload, the question to ask of any endpoint is not "what's the portal price today" but "what is the actual per-token list price the provider charges, and does anything sit between it and me." At $0.295/M output, Tencent Hy-MT2-7B just made that question interesting.

What to watch next

Three things follow from this week. First, the 1.8B and 30B-A3B siblings are now equally callable, so the "which size" decision is a real API question rather than a self-hosting decision (the family has its own matchup pages, but the short version is: 1.8B for on-device and the cheapest tier, 30B-A3B for professional-grade domains, 7B in between). Second, Tencent's WMT26 partnership offers awards to teams using Hy-MT models in the General Machine Translation and Video Subtitle Translation tasks — a signal that Tencent intends this family to be the open translation default in competitive MT. Third, the iOS and Android apps with on-device inference, if they ship as announced, will make the 1.8B the most installed open translation model in the world. The hosted API is the part that changed this week; the family's trajectory was set in May.