NewsBite analysis · AI, Cloud, Programming · New AI Model
The Decoder
Alibaba releases Qwen3.8-Flash-Next, targeting "ultimate cost efficiency"
Alibaba has introduced its latest AI model, Qwen3.8-Flash-Next, which aims to deliver superior performance with 125 billion parameters while only activating 6 billion per token. This model is tailored for high efficiency in specific tasks, resulting in significant cost savings compared to its predecessors. The newly designed N-gram embedding layer allows the model to utilize RAM efficiently, optimizing its performance in productivity and coding benchmarks.
Act nowConfidence: HighImmediate horizon
Published 26 Aug 2026, 16:10 UTC2 min read
Visit original article ↗Lead image — The Decoder
Summary
TL;DR
- Qwen3.8-Flash-Next features 125 billion total parameters, activating only 6 billion per token.
- The model promises better outcomes than Qwen3.7-Plus at one-ninth the training cost.
- It supports a context window of 262,144 tokens, scaling to 1 million tokens using YaRN technology.
- Produced as Qwen3.8-Flash, it is priced competitively at $0.16 per million input tokens.
- Flash-Next outperforms larger models like DeepSeek-V4-Flash and Anthropic's Claude Opus in many benchmarks.
Interesting facts
- The novel N-gram embedding layer functions as a phrase dictionary, enhancing efficiency.
- Despite fewer activated parameters, Flash-Next leads in coding and office task performance.