DEV.co
Open-Source LLM · tencent

Hy-MT2-30B-A3B

Hy-MT2-30B-A3B is a 30-billion parameter mixture-of-experts (MoE) multilingual translation model from Tencent, released May 2026. It supports 33 languages and instruction-following translation tasks (terminology, style, personalization, structured data). Apache 2.0 licensed, ungated, with 107k downloads. The model card claims outperformance vs. DeepSeek-V4-Pro and commercial APIs, but provides no independent benchmarks or security audit details.

Source: HuggingFace — huggingface.co/tencent/Hy-MT2-30B-A3B
30.1B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
107.9k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developertencent
Parameters30.1B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktranslation
Gated on HuggingFaceNo
Downloads107.9k
Likes469
Last updated2026-05-26
Sourcetencent/Hy-MT2-30B-A3B

What Hy-MT2-30B-A3B is

30B-A3B is a MoE-based causal LM trained for translation across 33 languages (zh, en, fr, pt, es, ja, tr, ru, ar, ko, th, it, de, vi, ms, id, tl, hi, pl, cs, nl, km, my, fa, gu). Supports transformers library (>=5.6.0). Distributed as safetensors. Context length unknown. Model card recommends temperature 0.7, top_p 1.0, top_k -1, repetition_penalty 1.0, max_tokens 4096. No quantized variants listed for 30B-A3B itself (1.8B has FP8, GGUF, 1.25-bit options). Inference code examples provided for transformers.

Quickstart

Run Hy-MT2-30B-A3B locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="tencent/Hy-MT2-30B-A3B")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

High-volume multilingual content localization

Supports 33 language pairs with instruction-following for terminology and style consistency, suitable for product, documentation, and marketing content.

Structured data and code comment translation

Model explicitly supports JSON, XML, and variable placeholder preservation, enabling safe translation of software artifacts and configuration files.

On-device or cost-constrained deployment

At 30B parameters, smaller than some alternatives; smaller variants (1.8B, 7B) available with extreme quantization (1.25-bit) down to 440 MB.

Running & fine-tuning it

ESTIMATE based on 30B MoE architecture: Single A100 80GB or multi-GPU setup (e.g., 2× A100 40GB) for full inference. Quantized (FP8/INT8) variants not listed for 30B; 16-bit inference likely requires 80GB+ VRAM. Verify exact MoE active parameter count and sparse activation pattern before deployment.

Card does not mention LoRA, QLoRA, or other fine-tuning compatibility. Given size (30B) and MoE architecture, LoRA is plausible but requires custom integration. Smaller variants (7B, 1.8B) more feasible for parameter-efficient tuning. Test on representative domain data before production.

When to avoid it — and what to weigh

  • Context-heavy or long-document translation — Context length unknown; card does not state max input length. Long documents may require chunking and context loss.
  • Real-time, ultra-low-latency requirements — 30B model size implies multi-GPU inference; no latency benchmarks provided. Smaller variants (1.8B) may be required for sub-second SLAs.
  • Specialized domain translation without fine-tuning — Card claims domain performance but does not detail which domains or provide domain-specific benchmarks. Custom fine-tuning may be necessary.
  • Regulatory/compliance audit trail required — No security audit, data provenance, or bias evaluation published. Suitable for internal/non-regulated use only without additional assessment.

License & commercial use

Apache License 2.0 (OSI-approved permissive license). Permits commercial use, modification, and distribution with attribution and liability disclaimer.

Apache 2.0 is a permissive OSI license that explicitly permits commercial use. However: (1) no indemnification clause; (2) card does not detail training data provenance or licensing (e.g., whether training corpus is commercially usable); (3) no warranty. Recommend legal review of training data sources and downstream liability before production deployment.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceMedium
Security considerations

No security audit, adversarial robustness testing, or bias evaluation disclosed. Translation models may be susceptible to prompt injection or adversarial inputs; treat outputs as machine-generated, not authoritative. Training data provenance unknown—no statement on data filtering, GDPR compliance, or PII handling. Recommend sandboxing untrusted model outputs in production.

Alternatives to consider

DeepSeek-V4-Pro (mentioned competitor)

Card claims Hy-MT2-7B and 30B-A3B outperform DeepSeek in fast-thinking mode. Independent eval required; DeepSeek is a general-purpose model, not translation-specific.

Google Translate API or Azure Translator

Established, compliant, and documented alternatives for production multilingual translation. Commercial SLA and support; suitable if latency/cost trade-off favors managed services.

NLLB-200 (Meta, open-source)

Smaller, mature, 200-language multilingual model. No MoE; lower inference cost. Fewer instruction-following features but well-documented and battle-tested.

Software development agency

Ship Hy-MT2-30B-A3B with senior software developers

Evaluate Hy-MT2-30B-A3B for your content and code translation pipeline. Review the model card, architecture report, and hardware requirements. Start with the smaller 1.8B or 7B variant for feasibility testing, then scale to 30B-A3B if latency permits.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Hy-MT2-30B-A3B FAQ

Can I use this model commercially?
Apache 2.0 permits commercial use with attribution and liability disclaimer. However, verify training data licensing separately—card does not disclose data sources or restrictions. Recommend legal review before production deployment.
What is the minimum hardware for inference?
Estimated: 80GB VRAM for single-GPU 16-bit inference on 30B-A3B; multi-GPU (2× A100 40GB) is practical. 1.8B and 7B variants support smaller GPUs (24GB–40GB). Use quantized variants (FP8, 1.25-bit for 1.8B) to reduce VRAM. Test on your hardware before production.
What is the context length?
Not stated in the model card. Recommended max_tokens is 4096, but card does not clarify total input + output context window. Verify by loading tokenizer or contacting Tencent.
Is fine-tuning supported?
Not documented. LoRA/QLoRA feasibility is unknown for MoE architecture. Recommend testing parameter-efficient tuning on your domain data, starting with smaller variants (1.8B, 7B) before scaling to 30B-A3B.

Work with a software development agency

DEV.co helps companies turn open-source tools like Hy-MT2-30B-A3B into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Ready to integrate multilingual translation?

Evaluate Hy-MT2-30B-A3B for your content and code translation pipeline. Review the model card, architecture report, and hardware requirements. Start with the smaller 1.8B or 7B variant for feasibility testing, then scale to 30B-A3B if latency permits.