Alibaba has released Qwen3.8-Flash-Next, a multimodal mixture-of-experts model designed for ultimate cost efficiency. The model is an architectural preview of Qwen4 and aims to deliver performance similar to much larger models while significantly reducing training costs. It has 125 billion total parameters but activates only 6 billion per token, with an additional 51 billion parameters in a novel N-gram embedding layer that operates in system RAM rather than GPU memory. The model supports a 262,144-token context window and can scale to one million tokens using YaRN. The technical report is available on GitHub, and weights are accessible on Hugging Face and ModelScope. The production version, Qwen3.8-Flash, is priced at $0.16 per million input tokens and $0.47 per million output tokens, with the API expected to launch soon. Source: thedecoder
Qwen3.8-Flash-Next outperforms Qwen3.7-Plus at roughly one-ninth the training cost, with significant improvements in coding and office tasks. The model scores 58.7 on DeepSWE and 62.5 on SWE-bench Pro, surpassing DeepSeek-V4-Flash and Claude Opus 4.6. It also achieves 73.9 on CoWorkBench, nearly double Qwen3.7-Plus's 65.1, and scores 55.7 on JobBench, nearly double its previous version's 27.6. On scientific reasoning and competitive programming tasks, the model is closely matched with other top models. Source: thedecoder
Alibaba's pricing strategy for Qwen3.8-Flash-Next puts pressure on competitors like OpenAI and Anthropic. The model costs about one-twelfth as much as Qwen3.8-Max, with a 12x price gap on both input and output tokens. Qwen3.8-27B, another model in the series, has gained popularity for its affordability and strong performance when run locally. OpenAI recently introduced discounts on its GPT-5.6 model line, which benefits users but may hinder rapid revenue growth for AI providers. Source: thedecoder
