[26/02/09] We have released HY-1.8B-2Bit, 2bit on-device large language model.
[26/01/13] We have released v0.3. We support the training and deployment of Eagle3 for all-scale LLMs/VLMs/Audio models, as detailed in the guidance documentation. And We released Sherry, the hardware-efficient 1.25 bit quantization algorithm [Paper Comming soon] | [Code]🔥🔥🔥
For more detailed information, please refer to[AngelSlim]
🌟HY-1.8B-2Bit Key Features
Superior Model Capability HY-1.8B-2Bit is developed via Quantization-Aware Training (QAT) based on the Hunyuan-1.8B-Instruct backbone. By aggressively compressing the model to a 2-bit weight precision, we achieve a performance profile that remains highly competitive with PTQ-INT4 benchmarks. Across a multi-dimensional evaluation suite—encompassing mathematics, humanities, and programming—HY-1.8B-2Bit exhibits a marginal performance degradation of only 4% compared to its full-precision counterpart, demonstrating exceptional information retention despite the radical reduction in bit-width.
Unmatched Scale-to-Performance Efficiency When compared to dense models of equivalent size (e.g., 0.5B parameters), HY-1.8B-2Bit demonstrates a substantial competitive advantage, outperforming benchmarks by an average of 16% across core competencies. As a state-of-the-art (SOTA) solution for its parameter class, HY-1.8B-2Bit provides an extensible and highly efficient alternative for edge computing, delivering high-tier reasoning capabilities within a compact footprint.
Comprehensive Reasoning Proficiency HY-1.8B-2Bit inherits the complete "full-thinking" capabilities of the Hunyuan-1.8B-Instruct model, marking it as the industry's most compact model to support sophisticated reasoning pathways. By integrating a Dual Chain-of-Thought (Dual-CoT) strategy, the model empowers users to navigate the trade-off between latency and depth: utilizing concise short-CoT for intuitive queries and detailed long-CoT for computationally intensive tasks. This flexibility ensures that HY-1.8B-2Bit can be seamlessly deployed in real-time, resource-constrained environments that demand both rapid response and high-fidelity logical synthesis.
📈 Benchmark
Benchmark results for HY-1.8B-2Bit equivalent weights on vLLM across cmmlu,ceval,arc,bbh,gsm8k,humaneval,livecodebench and gpqa_diamond.
The empirical results reveal that HY-1.8B-2Bit maintains high-tier performance despite the extreme reduction in bit-width, incurring a marginal average degradation of only 3.97% compared to its full-precision 1.8B teacher. Remarkably, HY-1.8B-2Bit performs nearly on par with the INT4 variant,with a negligible accuracy gap of only 0.13%, while utilizing only half the weight precision. When compared to the dense HY-0.5B model, which occupies a comparable model size, the superiority of the 2-bit QAT approach becomes evident. While the 0.5B dense model suffers a catastrophic 21.87% drop in average accuracy, HY-1.8B-2Bit remains robust, outperforming the smaller dense counterpart by 22.29% in GSM8K and 20.62% in LiveCodeBench.
Model
cmmlu
ceval
arc
bbh
gsm8k
humaneval (pass@3)
livecodebench
gpqa_diamond (pass@3)
HY-1.8B
55.07%
54.27%
70.50%
79.08%
84.08%
94.51%
31.50%
68.18%
HY-0.5B
37.08%
35.98%
49.89%
58.10%
55.04%
67.07%
12.11%
46.97%
HY-1.8B-int4gptq
50.80%
48.67%
68.83%
74.80%
78.70%
89.02%
30.08%
65.56%
HY-1.8B-2Bit
49.32%
47.60%
64.45%
75.54%
77.33%
93.29%
32.73%
65.15%
💻Deployment
This setup ONLY works on SME2-capable devices (for example, Apple M4, vivo x300 and Arm CPUs with SME2 support). Neon kernel will follow up.
@article{angelslim2026,
title={AngelSlim: A more accessible, comprehensive, and efficient toolkit for large model compression},
author={Hunyuan AI Infra Team},
journal={arXiv preprint arXiv:2602.21233},
year={2026}
}
💬 Technical Discussion
AngelSlim is continuously iterating and new features will be released soon. If you have any questions or suggestions, please open an issue on GitHub Issues or join our WeChat discussion group.
