Qwen3.8 Flash

This IntuitionLabs technical report covers Qwen3.8-Flash-Next architecture in depth, confirming 125B parameters with 6B activated per token, a 51B n-gram auxiliary table, and a 4B multi-token-prediction module, totaling approximately 180B parameters across 48 layers with native 262,144 context extensible to 1M via YaRN The article addresses deployment memory planning specifically for Qwen3.8-Flash-Next, noting the official FP8 checkpoint is 172.78 GiB (about 185.5 GB) of weight storage rather than full GPU-serving memory, with the vLLM recipe validating FP8 at TP2 minimum on GB300 and recommending TP4, and n-gram embedding requiring

This candidate explicitly names the hosted API model string "qwen/qwen3.8-flash" and ties it to Alibaba's Qwen 3.8 family, describing Qwen 3.8 Flash as a sparse mixture-of-experts multimodal reasoning model with 125B total parameters, 6B active per token, 512 experts, native 262,144-token context extensible to 1M, and The article provides benchmark and capability context specific to Qwen 3.8 Flash, positioning it as the cheap tier of the Qwen 3.8 family suited for product builds, with sibling context distinguishing it from the 27B local-rig variant and the Max variant for budget-unconstrained use. It reports modality support for tex

This buildfastwithai.com review directly covers the Qwen3.8-Flash-Next open-weight checkpoint, which is the foundation of the hosted Qwen3.8 Flash model, reporting that Qwen released weights on August 26, 2026 and that the architecture combines a 125B main model with 51B n-gram embedding parameters, 6B activated per to The article reports Qwen-published benchmark numbers including 91.9 on LiveCodeBench, 62.5 on SWE-bench Pro, 81.0 on SWE-bench Multilingual, 58.7 on DeepSWE 1.1, 73.9 on CoWorkBench, and 55.7 on JobBench, flagging these as provider-reported rather than independently verified. It positions the release as strategically i

This DataCamp article covers Qwen3.8-Flash-Next as an open-weight 125B MoE model with 6B active per token, previewing the Qwen4 architecture, released August 26, 2026, the same day as GLM-5.3 Flash. It explicitly compares the model to Claude Opus 4.6 Max on SWE-bench Pro (62.5 vs 53.4), CoWorkBench (73.9 vs 68.2), and The article frames Qwen3.8-Flash-Next as a cost-efficient workhorse below the Max-class models in raw intelligence but undercutting nearly everything on price, mirroring the Qwen3-Next-to-Qwen3.5 release pattern. Technical details include multimodal support, an experimental Qwen4 architecture preview status, and positi

This llm-stats.com aggregator leaderboard page covers the hosted Qwen3.8 Flash variant explicitly, reporting a composite LLM Stats Score of 48.7 at a blended price of $0.17 per 1M tokens, ranking it 24th overall with cost-efficiency positioning between DeepSeek-V4-Flash-0731 ($0.066) and DeepSeek-V4.1-Flash ($0.24). It The leaderboard data is sourced from the model's Hugging Face card and self-reported provider scorecards, and the page is flagged for captcha verification which can impede direct access, limiting confidence. Pricing appears to be a blended or reseller figure distinct from the QwenCloud direct rate of $0.16 per 1M input

Reuters reports that Alibaba's Qwen team released the multimodal Qwen3.8-Flash model on August 26, 2026, positioning it as a cost-efficient upgrade over its predecessor. According to Qwen's statement published on social media, the model supports a default context window of 262,144 tokens, expandable to 1 million tokens The article notes that Qwen would charge 1 yuan (approximately $0.1488) per million input tokens and 3 yuan per million output tokens for access through its application programming interface. In addition, Qwen released open-source weights for a separate Qwen3.8-Flash-Next variant, allowing the developer community to ev