Alibaba Cloud has launched Qwen2.5-Max, a large-scale Combination-of-Skilled (MoE) language mannequin, and made its API accessible by their cloud platform. The mannequin outperforms DeepSeek V3, which was launched final yr and made headlines for its coaching funds of $5.5 million.
“As we speak marks the Chinese language New 12 months, and whereas fireworks mild up the sky outdoors, right here I’m, sitting in entrance of my pc, penning this publish. We’ve lastly launched Qwen2.5-Max, an MoE mannequin on par with Deepseek-V3, now accessible on Qwen Chat and through API,” mentioned Binyuan Hui, a researcher at Alibaba Qwen Group.
The mannequin has been pretrained on 20 trillion tokens and additional refined utilizing Supervised Nice-Tuning (SFT) and Reinforcement Studying from Human Suggestions (RLHF) methodologies.
The mannequin’s efficiency was evaluated in opposition to main proprietary and open-weight fashions throughout varied benchmarks, together with MMLU-Professional, LiveCodeBench, LiveBench, and Enviornment-Exhausting. These benchmarks assess data, coding capabilities, normal skills, and human preferences, respectively.
“Qwen2.5-Max outperforms DeepSeek V3 in benchmarks corresponding to Enviornment-Exhausting, LiveBench, LiveCodeBench, and GPQA-Diamond, whereas additionally demonstrating aggressive ends in different assessments, together with MMLU-Professional,” the corporate said in its weblog publish.
The mannequin’s API, named qwen-max-2025-01-25, is now accessible on Alibaba Cloud. Customers can entry it by registering for an Alibaba Cloud account and activating the Mannequin Studio service. The API is suitable with OpenAI’s API, which makes it simple for builders to combine the mannequin into their functions.
Trying forward, the Qwen Group goals to additional improve Qwen2.5-Max’s capabilities by superior post-training methods.
“We’re devoted to enhancing the pondering and reasoning capabilities of enormous language fashions by the modern software of scaled reinforcement studying. This endeavour holds the promise of enabling our fashions to transcend human intelligence, unlocking the potential to discover uncharted territories of information and understanding,” the corporate mentioned.
Alibaba lately launched its newest vision-language mannequin, Qwen2.5-VL, which succeeds Qwen2-VL. This mannequin is constructed to “perceive issues visually,” together with recognising objects, analysing texts, charts, and graphics inside photographs, and performing as a visible agent able to directing instruments.
Considered one of its key options is that the mannequin can even management cell and pc screens, much like Anthropic’s Pc Use and OpenAI’s Operator agent.
The publish Alibaba Releases Qwen2.5-Max On Chinese language New 12 months, Outperforms DeepSeek V3 appeared first on Analytics India Journal.