Qwen3.8 Flash

DataCamp reports that Alibaba's Qwen team released Qwen3.8-Flash-Next on August 26, 2026, as an open-weight 125B mixture-of-experts model (6B active per token) serving as an early preview of the Qwen4 architecture. The model natively supports a 262,144-token context window, extensible to 1,000,000 tokens with YaRN, and On benchmarks, Qwen3.8-Flash-Next beats Claude Opus 4.6 Max on SWE-bench Pro (62.5 vs 53.4), CoWorkBench (73.9 vs 68.2), and JobBench (55.7 vs 36.6), while trailing on Humanity's Last Exam (35.9 vs 40.0). The article confirms the managed production "Qwen3.8-Flash" API is priced at $0.16 per million input tokens and $0.

Alibaba's Qwen team open-sourced Qwen3.8-Flash on August 26, 2026. The model is a multimodal Mixture-of-Experts design with 125B total parameters plus 51B n-gram embeddings, activating only 6B parameters per token. Weights were published on Hugging Face and ModelScope with day-0 vLLM and SGLang support, while QwenCloud hosts an API at $0.16 per 1M input tokens and $0.47 per 1M output tokens, positioning it for high-volume coding and agentic workloads. The release posts strong benchmark scores for its size class: 62.5 on SWE-bench Pro, 58.7 on DeepSWE 1.1, and 73.9 on CoWorkBench, beating Qwen3.7-Plus especially on coding and office-work tasks. Architecture highlights include GDN+QSA hybrid attention with up to 7.6x faster 1M-token prefill, Gated Residual, n-gram embeddings, and the Muon optimizer. Training reportedly cost roughly one-ninth that of Qwen3.7-Plus, and the model previews the Qwen4 series architecture.

LLM Stats tracks Qwen3.8 Flash as a managed production model with a blended price of approximately $0.17 per million tokens, positioning it competitively in cost efficiency against alternatives like DeepSeek-V4-Flash-0731 ($0.066) and GPT-6 Astra ($11.9). The tracker logs benchmark performance including MathVision (0.9 The tracker shows Qwen3.8 Flash with an LLM Stats Score of 49.1, placing it in a mid-tier cost-efficiency band relative to peers tracked on the platform. A quality-tracker time series displays performance variation across early September 2026, with the model holding a +0.40σ improving trend at the 7-day window. This in

Reuters reports that Alibaba's Qwen team released Qwen3.8-Flash on August 26, 2026, as a multimodal model delivering stronger coding and office-task performance while cutting training costs. The model supports a default context window of 262,144 tokens, expandable to 1 million tokens for handling large files, lengthy c Qwen announced API pricing of 1 yuan (~$0.1488) per million input tokens and 3 yuan per million output tokens for managed access. Alongside the managed release, Alibaba also published open-source weights for Qwen3.8-Flash-Next, a variant positioned as a prototype for the next-generation Qwen4 model family, enabling dev

Alibaba Cloud officially announced Qwen3.8-Flash as an open-weight, multimodal MoE model designed for optimal price-performance. It pairs a 125B-parameter main model with 51B N-gram embeddings and activates 6B parameters per token, natively supporting 262K tokens of context extensible to 1 million. Weights are downloadable from Hugging Face and ModelScope, with API access via Model Studio and Qwen Cloud at 1 RMB input / 3 RMB output per 1M tokens. Alibaba positions Qwen3.8-Flash as competitive against DeepSeek-V4-Flash and Claude-Opus-4.6 across benchmarks including SWE-bench Pro, CoWorkBench, Toolathlon Verified, MathVision, AndroidWorld, and ERQA, spanning agentic coding, long-horizon office tasks, multimodal reasoning, and embodied intelligence. The model also powers a redesigned Standard mode in QwenWork that cuts token consumption per task by 75% and roughly doubles generation speed. It serves as an early architectural preview of the upcoming Qwen4 series.

CellCog's tracker confirms Qwen3.8-Flash-Next shipped on August 26, 2026, with full weights and an FP8 variant on Hugging Face and ModelScope, alongside an official blog post and technical report on GitHub. It is described as an experimental preview of the Qwen4 architecture, with 125B main parameters, 6B activated per The article explicitly distinguishes the managed production version, stating "Qwen3.8-Flash" is priced at $0.16 per million input tokens and $0.47 per million output on QwenCloud, with the API now live as "qwen3.8-flash" and listed in Alibaba Cloud's Model Studio docs as of August 27, 2026. Day-one framework support ex

Vercel announced that Qwen 3.8 Flash from Alibaba is now available on its AI Gateway, taking text and images as input, serving a context window of 1 million tokens, and returning up to 65k tokens per response. Alibaba recommends the model for coding, tool use, and multi-step agent workflows. Developers can invoke it by Vercel's AI Gateway provides a unified API with usage tracking, retry, failover, custom reporting, Zero Data Retention support, and budgets. The gateway reflects provider pricing with no markup and does not charge a platform fee on inference, including on Bring Your Own Key requests. This integration post independently

OpenRouter lists qwen3.8-flash as a multimodal reasoning model from Alibaba, suited for coding assistance, agentic workflows, visual understanding, document and codebase analysis, desktop interaction, chart analysis, and long-video analysis. The listing confirms a release date of August 26, 2026, a 1M context window, a OpenRouter's telemetry reports throughput of 53 tokens per second (P50) and P50 latency of 3.26 seconds for Alibaba Cloud International, along with an average cache hit rate of 89.36%. The weighted-average effective input price is $0.03325 per 1M tokens, reflecting heavy caching, while weighted-average effective output