Release of Qwen 3.8 Model Series
The landscape of open-weight large language models (LLMs) has seen a significant expansion with the release of the Qwen 3.8 series. The release includes several iterations, most notably the Qwen 3.8 Max—officially identified in technical documentation as Qwen3.8-2.4T-A95B—and a 27B parameter model. While the API versions of these models reportedly offer multimodal capabilities, users on community forums such as Reddit have noted that the open-weight versions currently appear to be text-only, lacking integrated vision encoders.
Architectural Specifications and Performance
Technical discussions among developers have focused on the underlying architecture of the new models. There is ongoing speculation regarding whether the 27B model utilizes Multi-Token Prediction (MTP) or a different mechanism. Some users have noted that previous iterations, such as Qwen 3.5, shipped with an MTP layer out of the box, leading to expectations that the 3.8 architecture may follow suit.
In contrast to MTP, some developers have discussed the potential use of “DFlash,” a technique used in other models like Muse Glimmer to accelerate generation through speculative decoding. DFlash utilizes a lightweight “drafter” model to propose blocks of tokens, which the main model then verifies in parallel. However, it remains unverified whether the Qwen 3.8 27B model incorporates this specific architecture or relies on standard token-by-token generation.
Performance benchmarks shared by early testers suggest high efficiency, though specific hardware requirements remain a point of discussion. For instance, users have reported running previous versions at approximately 8 tokens per second on hardware with 32GB of unified memory.
Strategic Focus on Text-Based Models
The decision to release text-centric open-weight models, such as the 3.8 Max, rather than multimodal versions, has drawn scrutiny from the developer community. Some observers suggest this may be a strategic choice to prioritize foundational text capabilities over the higher computational costs associated with vision training. Others have speculated that the rapid release cycle—potentially following the release of competitors like Kimi K3—may have prioritized text-based performance over the integration of vision adaptors.
As the industry awaits further updates, including potential releases from other major players like Grok, the Qwen 3.8 series represents a significant addition to the available open-weight ecosystem, particularly for users seeking high-parameter text models for specialized tasks.