Qwen3.8-Flash-Next: 125 billion parameters, 6 active
Apache 2.0 weights, a 262,144-token context and $0.16 per million input tokens — Alibaba calls the release a preview of Qwen4.
Illustrative image: a narrow, blue-lit aisle running between long rows of server racks in a darkened data center.
Qwen3.8-Flash-Next is Alibaba's new multimodal mixture-of-experts model, released under Apache 2.0 with 125 billion parameters of which only 6 billion are active per token, and the company presents it as an architectural preview of Qwen4.
At a glance
- 125 billion total parameters, with 6 billion active per token (mixture of experts).
- Apache 2.0 license; weights and technical report on Hugging Face and ModelScope.
- Context window of 262,144 tokens natively, extendable to 1 million tokens with YaRN.
- QwenCloud pricing: $0.16 per million input tokens, $0.47 per million output tokens.
- Vendor claim: roughly one-ninth the training cost of Qwen3.7-Plus.
Alibaba released Qwen3.8-Flash-Next on August 26, 2026 — a multimodal mixture-of-experts model holding 125 billion parameters but running only 6 billion of them per token. The weights ship under Apache 2.0 via Hugging Face and ModelScope, alongside a technical report. A hosted production variant, Qwen3.8-Flash, runs on QwenCloud.
The n-gram layer that lives in system RAM
The headline design change is an n-gram embedding layer carrying 51 billion parameters. It stores frequently occurring word groups as standalone entries, which the-decoder likens to a phrase dictionary. The practical payoff is placement: that layer can sit in ordinary system memory rather than in expensive GPU memory.
According to the-decoder, this is the specific component Alibaba intends to carry into Qwen4. Flash-Next therefore reads less like a finished product line and more like a public trial run for the next generation.
Context and access
Native context is 262,144 tokens, stretchable to 1 million tokens using YaRN. Paired with Apache 2.0 terms, that puts commercial self-hosting on the table for anyone with the hardware to run it.
Benchmarks, as reported by the vendor
The scores below come from Alibaba's own report. None of the three sources reviewed here contains an independent re-run.
| Benchmark | Qwen3.8-Flash-Next |
|---|---|
| DeepSWE 1.1 | 58.7 |
| SWE-bench Pro | 62.5 |
| CoWorkBench | 73.9 |
| JobBench | 55.7 |
| GPQA Diamond | 91.7 |
| LiveCodeBench v6 | 91.9 |
Price is the real pitch
On QwenCloud, Flash-Next costs $0.16 per million input tokens and $0.47 per million output tokens. Alibaba's flagship Qwen3.8-Max sits at $2.00 input and $6.00 output. the-decoder.de puts the gap at roughly 1:12.
Training economics point the same way: Alibaba says Flash-Next required about one-ninth the training cost of Qwen3.7-Plus while scoring better. That figure, too, is unaudited. The company frames the whole release around what it calls ultimate cost efficiency.
One early hands-on account
Developer Simon Willison says he tried the model on a DGX Spark, using quantized builds of 72.5 GB and 78.9 GB. His write-up is a short link post centered on image outputs, not a structured evaluation. No official minimum memory requirement appears in any of the three sources.
What is still unknown
Alibaba has not given a Qwen4 release date, has not detailed the training data, and has not shown how the n-gram layer affects throughput at long context lengths. Verifying the benchmark table will take independent runs. For now the license and the prices are documented; the performance claims are not yet confirmed by anyone outside Alibaba.
FAQ
What is Qwen3.8-Flash-Next?
It is Alibaba's multimodal mixture-of-experts model with 125 billion parameters, 6 billion of which activate per token. It shipped on August 26, 2026 under Apache 2.0, and Alibaba describes it as an architectural preview of Qwen4.
How much does Qwen3.8-Flash-Next cost?
QwenCloud charges $0.16 per million input tokens and $0.47 per million output tokens, against $2.00 and $6.00 for Qwen3.8-Max. The weights themselves are free to download and use commercially under Apache 2.0.
Can you run Qwen3.8-Flash-Next locally?
No official hardware requirement is stated in these sources. Simon Willison reports running quantized builds of 72.5 GB and 78.9 GB on a DGX Spark, and the 51 billion parameters of the n-gram layer can be held in system RAM rather than GPU memory.
