Alibaba's Qwen3.8-Flash-Next slashes training costs with 125B params, active 6B, undercutting rivals
alibaba releases qwen3.8-flash-next, a 125b parameter multimodal model with 6b active params, at 1/12th the cost of qwen3.8-max, outperforming larger models.
alibaba's qwen team has released qwen3.8-flash-next, a multimodal mixture-of-experts model with 125b total parameters and 6b active params per token. it features a novel n-gram embedding layer with 51b parameters, stored in system ram. qwen3.8-flash-next supports a 262,144-token context window and scales to 1 million tokens using yarn. the model delivers better results than qwen3.7-plus at 1/9th training cost, with significant gains in coding and office tasks. qwen3.8-flash-next outperforms larger models like deepseek-v4-flash and claude opus 4.6 in most tasks. the production version, qwen3.8-flash, is priced at $0.16 per million input tokens and $0.47 per million output tokens, 12x cheaper than qwen3.8-max.