Alibaba's Qwen Team Announces More Open-Source Models Coming at ModelScope DevCon

At the ModelScope DevCon event in Nanjing, the team behind Alibaba's Qwen large language models made a brief but significant announcement via social media: "The answer everyone is waiting for is here: there will be more open Qwen models!"

The announcement, retweeted by developer Simon Willison, came directly from the official ModelScope account (@ModelScope2022). ModelScope is Alibaba's open-source AI model community and platform, which hosts the Qwen family alongside thousands of other models.

While the tweet provided no specific details on model sizes, capabilities, architecture, or release timelines, the commitment is clear: the Qwen series will continue to expand with new open-source releases.

Context: The Qwen Series Landscape

The Qwen (通义千问) series from Alibaba's Qwen team has established itself as a major contender in the open-weight LLM arena, competing directly with offerings from Meta (Llama), Mistral AI, and Google. Recent notable releases include:

• Qwen2.5: The latest major iteration, featuring model sizes from 0.5B to 72B parameters, with strong performance on benchmarks like MMLU, GPQA, and GSM8K.
• Qwen2.5-Coder: A specialized model for code generation.
• Qwen2-VL and Qwen2-Audio: Multimodal variants for vision and audio tasks.

All Qwen models are released under the permissive Apache 2.0 license, allowing for commercial use, modification, and distribution—a key factor in their adoption by developers and enterprises.

What "More Open Models" Could Mean

Based on the trajectory of the Qwen team and competitive pressures, the new models could take several forms:

1. Larger-Scale Models: A push beyond the current 72B parameter flagship to compete with the largest open models, potentially targeting a 100B+ parameter class.
2. Specialized Variants: Further domain-specific models, perhaps targeting mathematics, scientific reasoning, or long-context processing.
3. Efficiency-Focused Models: Smaller, highly optimized models for edge deployment or faster inference.
4. Multimodal Expansion: Enhanced or new versions of their vision-language (VL) or audio models.

The announcement serves primarily as a strategic signal to the developer community that Alibaba remains committed to the open-source model ecosystem, encouraging continued adoption and integration of the Qwen platform.

gentic.news Analysis

This announcement, while light on specifics, is a calculated move in the high-stakes open LLM war. The Qwen team is likely battling two pressures: the relentless release cadence of competitors like Meta and Mistral AI, and the internal need to demonstrate continued ROI on their substantial R&D investment to Alibaba's leadership. A vague "more is coming" tweet costs nothing but maintains mindshare and reassures developers building on Qwen that the pipeline isn't dry.

Technically, the most anticipated gap in their lineup is a true top-tier, 100B+ parameter model that can go head-to-head with the best of Llama 3.1 405B or Google's Gemma 2 27B in efficiency-per-parameter. The Qwen2.5 72B model is highly capable, but the frontier is pushing toward larger scales. Alternatively, they may be focusing on a mixture-of-experts (MoE) architecture, which has become a standard tool for achieving high capability with lower inference cost, as seen in Mistral's Mixtral and Meta's Llama 3.1 405B. Their silence on details suggests the team is still finalizing architecture decisions or training runs.

For practitioners, the key takeaway is that the open-source LLM market remains fiercely competitive, which is excellent for end-users. However, it also creates a form of version fatigue. The promise of "more models" is less compelling than it was a year ago. What developers need now are not just more models, but more stable, long-term supported model families with clear upgrade paths, robust tooling, and predictable release cycles. The Qwen team's next challenge is to transition from a series of impressive releases to becoming a dependable platform.

Frequently Asked Questions

What is the Qwen model series?

Qwen is a family of large language models developed by Alibaba Group. The models range from 0.5 billion to 72 billion parameters, are trained on multilingual data, and are released under the open-source Apache 2.0 license. They are known for strong performance in reasoning, coding, and multilingual tasks.

What is ModelScope?

ModelScope is an open-source AI model community and platform launched by Alibaba. It hosts thousands of models, including the entire Qwen family, and provides tools for model discovery, experimentation, and deployment. It is Alibaba's primary channel for distributing its AI research to the public.

When will the new Qwen models be released?

The announcement did not provide any release dates or timelines. The tweet only confirmed that more models are in development. Release schedules are typically announced via the official ModelScope platform and GitHub repository when models are ready for launch.

How do Qwen models compare to Llama or Mistral models?

The Qwen series is a direct competitor to Meta's Llama and Mistral AI's models. Benchmarks show competitive performance, with Qwen2.5-72B often matching or exceeding comparable models like Llama 3 70B on reasoning and coding tasks. A key differentiator is Qwen's strong multilingual capability, with particularly good performance on Chinese language tasks.