What is the N-gram embedding layer?

It is a novel layer with 51 billion parameters that stores common word groups as standalone entries. This 'phrase dictionary' runs in system RAM rather than GPU memory at relatively low additional cost.

How does Qwen3.8-Flash-Next compare to Qwen3.7-Plus?

It delivers better results than Qwen3.7-Plus at roughly one-ninth the training cost. Qwen3.7-Plus has 397 billion parameters with 17 billion activated per token.

What is the pricing for the production version?

The production version, Qwen3.8-Flash, is priced at $0.16 per million input tokens and $0.47 per million output tokens through QwenCloud.

Large Language ModelsOpen-Source AIAI Business & IndustryTHE DECODERAug 26, 2026Aug 27, 2026, 01:01 JST

Alibaba's Qwen3.8-Flash-Next Targets Cost Efficiency

Send on LINE

3 Key Points

What happened
Alibaba's Qwen team released Qwen3.8-Flash-Next, a multimodal mixture-of-experts model. It has 125 billion total parameters but only activates 6 billion per token.
Why it matters
The model delivers better results than Qwen3.7-Plus at roughly one-ninth the training cost, according to the Qwen team. It leads the majority of tested tasks against much larger models like DeepSeek-V4-Flash and Claude Opus 4.6.
What to watch
The production version, Qwen3.8-Flash, is priced at $0.16 per million input tokens and $0.47 per million output tokens through QwenCloud. This is about one-twelfth the cost of the flagship Qwen3.8-Max and is seen as pressure on rivals like OpenAI and Anthropic.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The release of Qwen3.8-Flash-Next is positioned as an architecture preview for the upcoming Qwen4. A key innovation is the N-gram embedding layer, which stores common word groups in a 'phrase dictionary' that runs in system RAM, not GPU memory. This approach contributes to the model's cost efficiency, with a total of 125 billion parameters but only 6 billion activated per token.

The model's performance is notable in agentic coding and office tasks. It scored 58.7 on DeepSWE and 62.5 on SWE-bench Pro, surpassing both DeepSeek-V4-Flash and Claude Opus 4.6. The gap is wider on office productivity, where it hit 73.9 on CoWorkBench compared to DeepSeek's 45.1, and 55.7 on JobBench. Claude Opus 4.6 only outperforms it on Humanity's Last Exam, which the article notes is an older model from February 2026.

The pricing strategy continues to apply pressure on competitors. Flash-Next costs about one-twelfth as much as the flagship Qwen3.8-Max, with a roughly 12x price gap on tokens. This keeps pressure on OpenAI and Anthropic, especially since OpenAI recently responded with discounts on its own GPT-5.6 models. This trend is good for users but potentially challenging for AI providers that need rapid revenue growth to sustain their investment narratives.

FAQ

What is the N-gram embedding layer?: It is a novel layer with 51 billion parameters that stores common word groups as standalone entries. This 'phrase dictionary' runs in system RAM rather than GPU memory at relatively low additional cost.
How does Qwen3.8-Flash-Next compare to Qwen3.7-Plus?: It delivers better results than Qwen3.7-Plus at roughly one-ninth the training cost. Qwen3.7-Plus has 397 billion parameters with 17 billion activated per token.
What is the pricing for the production version?: The production version, Qwen3.8-Flash, is priced at $0.16 per million input tokens and $0.47 per million output tokens through QwenCloud.

THE DECODERRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.