Alibaba has released Qwen3.8-Flash-Next, an open-weight multimodal model built on what the company describes as an entirely new architecture, and the benchmark numbers accompanying the release put it ahead of Claude Opus 4.6 (Max) on the majority of tests the two models were compared on. The model is available now through ModelScope, and Alibaba has priced it as a mid-tier release rather than a flagship, which makes the comparisons to Anthropic’s top-end model somewhat unusual on the surface.
Qwen3.8-Flash-Next combines a 125B-parameter Mixture-of-Experts backbone with an additional 51B parameters devoted to n-gram embeddings, but only 6B parameters are active for any given token. That activation ratio is the whole point of the release. Alibaba is pitching this as a model that delivers most of the reasoning and coding capability of its larger Qwen3.8 line while running at a fraction of the compute cost, and the company claims training costs came in at roughly one-ninth of what Qwen3.7-Plus required.
The context window is native 256K, extendable to 1 million tokens through YaRN, and Alibaba says the attention kernels behind that extended window have been reworked substantially. At the 1M token mark, the company is reporting prefill speeds up to 7.6 times faster and decode speeds up to 4.9 times faster than what the architecture would otherwise deliver, a claim that if it holds up in independent testing would matter quite a bit for anyone running long-document or long-codebase workloads where latency has historically been the bottleneck.
Qwen3.8-Flash-Next Benchmarks
On the benchmark table itself, the coding numbers are where Alibaba wants attention. Qwen3.8-Flash-Next scores 62.5 on SWE-bench Pro against Claude Opus 4.6 Max’s 53.4, and it also leads on SWE-bench Multilingual with 81.0 versus 77.5. DeepSeek-V4-Flash-0731 comes closest to matching Qwen on the agentic coding front, scoring 54.4 on DeepSWE 1.1 to Qwen’s 58.7, though it slips further behind on the multilingual test. Qwen3.7-Plus, despite being the larger and more expensive model in Alibaba’s own lineup, trails its newer sibling across every coding category in the table, which is presumably the comparison Alibaba cares most about making.
The agent category tells a similar story. Qwen3.8-Flash-Next posts a 73.9 on CoWorkBench, ahead of Claude Opus 4.6 Max’s 68.2, and a wide 55.7 on JobBench compared to 36.6 for the Anthropic model. It also leads on Toolathlon Verified at 73.5. The one place Claude pulls ahead in the agent section is Humanity’s Last Exam, where Claude Opus 4.6 Max’s 40.0 score beats Qwen’s 35.9, and DeepSeek edges out the field on Agents’ Last Exam Pass@1 with a 25.2 to Qwen’s 24.3. Frontier reasoning benchmarks of this kind have tended to be the one area where Anthropic and OpenAI’s top models hold a clearer edge over Chinese open-weight releases, and that pattern shows up again here.
Alibaba’s own Qwen3.7-Plus also appears in the table as a point of internal comparison, and it loses to Flash-Next almost everywhere despite presumably costing more to run. That framing is deliberate. Much of Alibaba’s messaging around this release leans on efficiency rather than raw scale, following the same direction the company took with Qwen3.8-27B a few weeks earlier, where a smaller model was positioned against considerably larger rivals and still came out competitive.
General capability scores round out the picture. Qwen3.8-Flash-Next posts a 91.7 on GPQA Diamond, just short of Claude’s 91.3, and hits 91.9 on LiveCodeBench v6 against Claude’s 88.8. Instruction following on IFBench is where the gap widens the most in Qwen’s favor, 81.3 versus 62.5, though Claude’s models have historically optimized for different priorities than raw instruction-benchmark scores, and Anthropic’s own releases have tended to score lower on IFBench even when the underlying model performs well in practice.
Whether these numbers translate into real adoption is a separate question from whether they’re accurate. Alibaba has released several Qwen models in quick succession this year, and each has arrived with a chart showing it ahead of some subset of Western frontier models on some subset of benchmarks. What makes Flash-Next notable is the size of the model doing the beating. A 6B-active-parameter model outperforming a frontier-class Opus release, even on a curated set of benchmarks Alibaba chose to publish, says something about how quickly the efficiency gap between open and closed models has been closing this year.