Performance Benchmarks for Qwen 3.8 Flash Next
New performance data for the Qwen 3.8 Flash Next model has surfaced, positioning the model within a competitive tier of large language models. According to updated scores from Artificial Analysis, Qwen 3.8 Flash Next has achieved a score of 29. This placement puts the model between GPT-6 Luna Max, which scored 28, and Claude Haiku 5.5 High, which scored 30.
In comparison to other models within the Qwen 3.8 family, the Flash Next iteration demonstrates significant capability gains. Data indicates the model is performing closer to the Qwen 3.8 Max (score of 34) than to the 27B version (score of 20). While some users on community forums have noted that the model may feel more capable than Claude Haiku 5.5 in terms of reasoning and “understanding,” others have observed that it may exhibit slower inference speeds.
Nace AI Introduces Weight-Space Quantization
The open-source AI lab Nace AI has announced the release of new GGUF weights for several Qwen models, including the 9B, 27B, 35B, and Flash Next versions. These weights were produced using a technique referred to as “Weight Space Programming” (WSP).
According to documentation provided by Nace AI, the quantization process utilized an “agentic swarm” of models—specifically Astra, Fable, and Carbon—to propose and refine tensor-precision changes. The goal of this method is to minimize Kullback–Leibler (KL) divergence, a metric used to measure how much a quantized model deviates from the original full-precision model.
Nace AI reports that their quantized versions achieved high top-1 accuracy across various subsets, including prose, chat, and code. Specifically, the Flash Next IQ4_XS quantization reached a top-1 accuracy of 93.0% at 93 GB. The 27B Q4_K_XL version reported a 97.7% accuracy at 17.3 GB.
Benchmarking and Long-Context Stability
In testing conducted by Nace AI, the new quantizations were evaluated for stability over long context windows, specifically up to 128K tokens. The lab claims that their models showed less “long-context drift” compared to existing alternatives. For instance, the 27B model showed minimal movement in KL divergence when moving from a 512-token context to a 128K-token context.
The lab also measured “p99 KL divergence” to identify catastrophic failures in the worst 1% of token predictions. Their results suggest that the Flash Next and 27B models maintained stable tails in their distribution, avoiding the significant errors sometimes seen in other quantization methods.
While Nace AI asserts that these quants hold the “frontier” for their respective sizes in terms of KL divergence, these results are based on the lab’s internal benchmarks and have not been independently verified by a third-party auditing firm. The lab has indicated that further research and a formal paper regarding the scalability of the Weight Space Programming method are forthcoming.
