Tencent puts Hy-MT2 translation on iPhone

Tencent has released Hy Translate, a standalone iOS app that translates text, speech, and images without sending each request to a cloud service. The app launches in 12 countries and regions with support for 33 languages plus five Chinese minority languages and dialects.

Local inference gives Tencent a practical way to compete with Google Translate, DeepL, and Microsoft Translator. It removes network latency and per-request API costs while allowing translation on planes, in subways, and anywhere else without a reliable connection.

The 440MB route to local inference

The open-source Hy-MT2 family includes three models aimed at different deployment budgets:

Tencent designed the 1.8B model for phones, although the App Store listing does not identify the exact checkpoint used by Hy Translate. The company says its AngelSlim 1.25-bit quantization compresses the mobile model to roughly 440MB and supports chipsets from Apple, Qualcomm, and MediaTek.

Quantization stores model weights at lower precision, reducing storage and memory use at the risk of some quality loss. Tencent reports inference performance 1.5 times faster than its previous generation, but its launch materials do not identify the comparison model, test device, or workload.

The 30B-A3B release uses a mixture-of-experts architecture that routes each token through a subset of specialized parameters. Only 3 billion parameters are active for a given token, which reduces computation compared with activating the full 30 billion. The 7B and 1.8B models are dense, so every parameter participates in each generation step.

Benchmarks come with boundaries

Tencent reports that the 7B and 30B-A3B models outperform DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode. In this context, fast-thinking means generating the translation directly instead of running an extended reasoning pass, which keeps latency lower.

The company also says the 1.8B model exceeds Microsoft and Doubao translation APIs on aggregate evaluations. Those claims apply to model-level tests rather than measured performance inside the iOS app. The published summary does not include per-language results, device-level throughput, peak memory use, battery consumption, or thermal behavior.

Tencent released IFMTBench alongside the models to test instruction-following translation. Its tasks cover constraints such as preserving JSON structure, retaining subtitle timing, and following formatting instructions. Aggregate benchmark scores can conceal weak language pairs or domain-specific failures, so production evaluations should use representative text from the intended workload.

The developer path spans servers and edge devices

The Hy-MT2 release includes weights and tooling across Hugging Face, ModelScope, and GitHub. Developers can use the models in several ways:

• Run the original checkpoints for evaluation or self-hosted translation.
• Fine-tune all parameters on a domain corpus.
• Use LoRA adapters to reduce fine-tuning memory and storage requirements.
• Train with supported DeepSpeed ZeRO configurations or LLaMA-Factory.
• Deploy quantized builds on supported ARM and Intel hardware.

Server-side evaluation can start with the documented SGLang launch pattern:

python3 -m sglang.launch_server \
  --model tencent/Hy-MT2-1.8B \
  --tp 1

This command starts a server using one tensor-parallel worker. Mobile deployment requires the separate quantized runtime and device-specific integration rather than the SGLang server path. Teams should also review the repository license before distributing a modified model or embedding it in a commercial product.

A standalone app tests the full product

Tencent previously exposed Hy-MT2 through a WeChat mini-program and said native iOS and Android apps would follow with local inference. The iOS release delivers the first of those native clients, while an Android launch remains pending.

Packaging the model in a consumer app tests more than translation quality. Tencent must manage model downloads, memory pressure, startup time, battery use, camera and microphone input, and updates across several generations of mobile hardware.

Local processing also narrows the amount of translation content that needs to reach a remote API. That architecture can reduce data exposure, although the app’s telemetry, crash reporting, and privacy policy still determine what other information leaves the device.

Small models reshape translation costs

Hy-MT2 demonstrates how a specialized 1.8B model can fit into a sub-500MB deployment package while competing with larger hosted systems on Tencent’s reported evaluations. For high-volume applications, local execution replaces per-call charges with device compute, model distribution, and maintenance costs.

Cloud services still offer centralized updates, elastic capacity, and access to larger models. On-device models add predictable latency, offline operation, and tighter control over translation inputs. The useful balance depends on language coverage, hardware limits, quality targets, and update requirements.

Tencent has also previewed Hy ASR 3.0 for speech recognition, indicating a broader strategy built around compact, task-specific models. Translation provides the first consumer deployment, with speech recognition and other narrow workloads offering similar opportunities for local inference.