Tencent's 440MB On-Device Translation: 1.25bit Quantization, 33 Languages, Beats 72B Models

On April 29, 2026, Tencent Hunyuan Team open-sourced Hy-MT1.5-1.8B-1.25bit on-device translation model through AngelSlim.

Key metrics: Model size only 440MB (87% compression rate), natively supports 33 languages and 1,056 translation directions, translation quality comprehensively surpasses large-parameter models like Tower-Plus-72B and Qwen3-32B, as well as mainstream commercial APIs like Microsoft Translator and Doubao Translator.

Technical breakthrough: Employs Sherry 1.25bit quantization (accepted at ACL 2026), using 3:4 fine-grained sparsity strategy to compress the original 3.3GB FP16 model to 440MB with minimal accuracy loss. Paired with custom STQ mobile CPU kernel, ordinary phones can run high-quality offline translation smoothly.

Model Versions: Two Quantization Options

Hy-MT1.5-1.8B provides two quantized versions:

Version	Model Size	Quantization	Compression	Use Cases
Hy-MT1.5-1.8B-1.25bit	440MB	1.25bit	87%	Memory-constrained devices, extreme size
Hy-MT1.5-1.8B-2bit	574MB	2bit	83%	Balance performance and size

Both versions support full 33 languages and 1,056 translation directions, with both weight files and GGUF format available.

Core Features

1. World-Class Translation Quality

Hy-MT1.5-1.8B-1.25bit is built upon the Hy-MT1.5-1.8B foundation model developed by Tencent Hunyuan Team.

Training Pipeline:

• MT-oriented pre-training
• Supervised fine-tuning (SFT)
• On-policy distillation
• Reinforcement learning (RL)

Language Coverage:

• Natively supports 33 languages
• Supports 5 dialects/minority languages
• Covers 1,056 translation directions

Performance Comparison:

Comparison	Hy-MT1.5-1.8B (1.8B)	Tower-Plus (72B)	Qwen3 (32B)
Parameters	1.8B	72B	32B
Translation Quality	✅ Comprehensively surpasses	❌ Behind	❌ Behind
Model Size (1.25bit)	440MB	-	-

With only 1.8B parameters, comprehensively surpasses 72B and 32B large-parameter models, while also beating mainstream commercial APIs like Microsoft Translator and Doubao Translator.

2. Sherry: Extreme 1.25bit Quantization

Most eye-catching is the Sherry quantization framework (accepted at ACL 2026).

Core Innovation: 3:4 Fine-grained Sparsity

Sherry introduces a hardware-efficient ternary quantization strategy:

• For every 4 model weights, keep the 3 most important weights, stored in 1-bit (1)
• Remaining 1 weight is zeroed out
• 4 weights require only 5 bits storage
• Achieves effective 1.25-bit width with power-of-two alignment

Compression Results:

• Original FP16 model: 3.3GB
• After 1.25bit quantization: 440MB
• Compression rate: 87%
• Accuracy loss: Minimal

Technical Advantages:

Technical Metric	Sherry 1.25bit	Traditional Quantization
Bit Width	1.25bit	Usually 4-8bit
Compression Rate	87%	50-75%
Hardware Alignment	✅ Power-of-two aligned	❌ May not align
SIMD Optimization	✅ Perfect alignment	❌ Lower efficiency
Accuracy Loss	Minimal	Moderate

3. On-Device Deployment: Runs on Ordinary Phones

Paired with custom STQ kernel (designed specifically for mobile CPUs), the 1.25bit model achieves perfect SIMD instruction set alignment.

Deployment Advantages:

• Ordinary phones (memory-constrained) can run smoothly
• No internet required
• Data never leaves device
• High-quality offline translation
• One-time download, permanent use

Speed Comparison:

Tested on Snapdragon 888, 8GB RAM device:

• FP16 model: Baseline speed
• 1.25bit model: 8x speedup

Translation Performance: Flores-200 Benchmark

On Flores-200 Chinese-Foreign mutual translation benchmark, Hy-MT1.5-1.8B demonstrates excellent performance:

Performance Comparison Across Model Sizes:

Model	Parameters	Translation Quality (BLEU)	Model Size (1.25bit)
Hy-MT1.5-1.8B	1.8B	✅ Highest	440MB
Tower-Plus	72B	❌ Lower	-
Qwen3	32B	❌ Lower	-

The data doesn’t lie: 1.8B parameter Hy-MT1.5 comprehensively surpasses 72B and 32B large-parameter models in translation quality.

Android Demo: Ready-to-Use Offline Translation

Tencent Hunyuan Team provides a ready-to-use Android Demo showcasing on-device translation capabilities.

Core Features:

1. Offline Translation

• Supports 33 language mutual translation
• No internet required
• No data upload
• One-time download, permanent use

2. Background Word Extraction Mode (Highlight Feature)

Cross-App Word Translation:

• While browsing emails, webpages, chat messages
• No app switching required
• Instant translation
• Completely offline

Use Cases:

• Reading foreign emails
• Browsing foreign webpages
• Viewing foreign chat messages
• Learning foreign language content

Test Devices:

• Translation Demo: Snapdragon 865, 8GB RAM
• Background Word Demo: Snapdragon 7+ Gen 2, 16GB RAM

Download Links

Model Weights

Version	Weight Files	GGUF Format
1.25bit	ModelScope	GGUF
2bit	ModelScope	GGUF

Android Demo

Download: https://modelscope.cn/models/AngelSlim/Hy-MT1.5-1.8B-1.25bit-GGUF/resolve/master/Hy-MT-demo.apk

Technical Reports

Report	Link
HY-MT1.5 Technical Report	https://arxiv.org/abs/2512.24092
Sherry Paper (ACL 2026)	https://arxiv.org/abs/2601.07892
AngelSlim Technical Report	https://arxiv.org/abs/2602.21233

AngelSlim: Tencent’s LLM Compression Toolkit

Hy-MT1.5-1.8B-1.25bit is one of the achievements from the AngelSlim toolkit.

AngelSlim Overview:

• Tencent’s self-developed large language model compression toolkit
• Dedicated to building more accessible, comprehensive, and efficient compression solutions

Open-Source Algorithms:

• Quantization
• Speculative Sampling
• Sparsification
• Distillation

Coverage:

• Mainstream cutting-edge large models
• End-to-end pipeline from compression to deployment

GitHub: https://github.com/Tencent/AngelSlim

Significance of Technical Breakthrough

1. New Benchmark for On-Device AI

Hy-MT1.5-1.8B-1.25bit proves that:

• On-device models can achieve world-class quality
• Extreme compression doesn’t mean performance compromise
• 440MB can support 33 languages and 1,056 translation directions

2. New Breakthrough in Quantization

Sherry 1.25bit quantization:

• Breaks traditional quantization bit-width limits
• Achieves hardware-efficient ternary quantization
• 87% compression rate + minimal accuracy loss

3. New Solution for Privacy Protection

Completely offline operation:

• No data upload
• No network dependency
• Complete privacy protection

Application Scenarios

Scenario	Core Capability	Practical Value
Offline Translation	33 language mutual translation	Translation needs without network
Cross-App Word Extraction	Background word translation	Instant translation while reading foreign content
Privacy Translation	Completely offline	Sensitive content translation, no data leakage
Learning Assistant	High-quality translation	Foreign language learning, literature reading
Business Communication	Multi-language support	Cross-border business communication, email translation

Final Thoughts

The release of Hy-MT1.5-1.8B-1.25bit marks a new phase for on-device AI.

It’s no longer about “can run on phones” to call it an on-device model, but achieving: 440MB size + 33 languages + world-class translation quality + completely offline operation.

More importantly, Tencent Hunyuan Team not only open-sourced the model weights but also provided:

• Complete technical reports (HY-MT1.5, Sherry, AngelSlim)
• Ready-to-use Android Demo
• GGUF format support
• Open-source compression toolkit (AngelSlim)

From technical innovation perspective, Sherry 1.25bit quantization (accepted at ACL 2026) achieved:

• 3:4 fine-grained sparsity
• Hardware-efficient ternary quantization
• 87% compression rate + minimal accuracy loss
• Perfect SIMD instruction set alignment

From practical application perspective, the 1.8B parameter model comprehensively surpasses 72B and 32B large-parameter models in translation quality, proving the “small but precise” technical approach is completely viable.

The future of on-device AI may come faster than we think.