Release of Qwen 3.8 Series Models

The Qwen model series has expanded with the release of the 3.8 iteration, including a high-parameter 2.4T-A95B version and a more accessible 27B parameter model. According to data shared on Hugging Face and discussed on community forums such as Reddit, the 2.4T-A95B model represents a significant scale in parameter count, while the 27B variant is being positioned as a high-performance option capable of running on consumer-grade hardware, such as an NVIDIA RTX 3090.

Early user reports on Reddit suggest that the 27B model demonstrates improved reasoning and judgment capabilities compared to its predecessors, Qwen 3.5 and 3.6. In one anecdotal comparison involving security vulnerability analysis, a user noted that the 3.8 model correctly identified a “false positive” in a code analysis pipeline—a nuance that the 3.6 model had previously flagged as a legitimate vulnerability. However, such comparisons remain unverified and are subject to variables such as quantization levels and hardware configurations.

Benchmark Performance and Capabilities

The Qwen 3.8-2.4T-A95B model has been evaluated against several industry-standard benchmarks. According to data published on Hugging Face, the model shows competitive performance in several categories:

  • Coding and Software Engineering: The model achieved a score of 84.6 on the Terminal Bench and 84.6 on the Coding Agent benchmark. On the SWE-bench Pro, it recorded a score of 69.2, while its performance on DeepSWE was noted at 59.0.
  • General Capabilities: In the GPQA Diamond benchmark, the model scored 92.0. It also recorded a 45.7 on the HLE benchmark and 62.2 on the IFBench.
  • Long Context and Reasoning: The model achieved a 69.1 on LongBench v2 and an 83.2 on the MRCR v2 (8-needle) test.

The 27B model has also been noted for its performance on the Artificial Analysis Benchmark, where it reportedly places in proximity to models such as DeepSeek V4 and GPT-5.6 Luna Max.

Technical Observations and Community Feedback

Technical discussions regarding the 27B model have focused on its efficiency and “thinking” processes. Some users have raised questions regarding whether the model suffers from excessive “overthinking” or high token consumption during reasoning tasks. One user reported a generation of 37,077 tokens taking approximately 14 minutes, highlighting the computational demands of extended reasoning.

In the realm of model safety and “abliteration,” unverified reports suggest that certain versions of the Qwen 3.8-27B FP8 model have seen a significant reduction in refusal rates—dropping from a base refusal rate of 64–99% to between 0–6%—while maintaining relatively stable scores in MMLU (84.7) and GSM8K (88.7). Analysts on community forums have cautioned that these “abliterated” versions are intended for red-teaming purposes and that the impact on long-horizon reasoning and tool-calling remains unconfirmed.

While the 2.4T-A95B model offers high-tier performance, users have noted the absence of vision capabilities in this specific release. Additionally, the community continues to await GGUF quantizations to facilitate easier local deployment of the larger models.