Qwen3.8 Flash

DataCamp reports that on the same day as Z.ai's GLM-5.3 Flash launch, Alibaba opened the weights of Qwen3.8-Flash-Next, a multimodal mixture-of-experts model that previews the upcoming Qwen4 architecture, mirroring how Qwen3-Next previewed Qwen3.5. The model is a 125B-parameter MoE with 6B active per token, supports a According to the article, Qwen3.8-Flash-Next is priced at $0.16 per million input tokens and $0.47 per million output tokens on QwenCloud under the qwen3.8-flash API SKU. Vendor benchmarks cited include SWE-bench Pro at 62.5 versus Claude Opus 4.6 Max at 53.4, CoWorkBench at 73.9 versus 68.2, and JobBench at 55.7 versu

Orca Router's coverage frames Alibaba's most useful August 26, 2026 release as the accompanying technical paper, "On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability," which publishes ablations, negative results, and training-recipe experiments tied to the Qwen4 family. The model The article lists the model as priced at $0.15 per 1M input tokens and $0.47 per 1M output tokens under the Qwen: Qwen3.8 Flash entry dated 2026-08-26, positioned in a release-timeline alongside GLM 5.3 Flash, DeepSeek V4 Flash Vision, and others. The piece references the official Alibaba technical report as the primar

BenchLM records Qwen3.8-Flash-Next as released August 26, 2026, open-weight with reasoning support and 262K context, ranking it at a composite capability score of 61.8/100 and 58th of 228 tracked models as of August 31, 2026. Category percentiles show Agentic at the 96th percentile (rank 7 of 142), Coding at the 82nd p The model record reports 87 tokens/sec output speed and a 25.42-second time-to-first-token, with self-hosted infrastructure costs flagged as not comparable to API token rates and no first-party hosted token rate published. BenchLM notes 24 source-displayable benchmark rows and highlights Agentic capability as the model

Qwen3.8-Flash-Next, released August 26, 2026, is the open-weight counterpart to Alibaba's production API SKU qwen3.8-flash and serves as the second-generation "Next" architecture preview ahead of the Qwen4 family. According to Local AI Zone's technical analysis, the model stores 180B parameters (125B main MoE backbone The article also reports vendor self-published benchmarks including DeepSWE 1.1 at 58.7, SWE-bench Pro at 62.5 (framed as beating Claude Opus 4.6 Max's 53.4), CoWorkBench at 73.9, GPQA Diamond at 91.7, LiveCodeBench v6 at 91.9, and AndroidWorld at 84.5, with weaker showings on NL2Repo-Bench (48.1 vs DeepSeek-V4-Flash 5

Alibaba's Qwen team released Qwen3.8 Flash, a new multimodal model, on August 26, 2026, positioning it with stronger coding and office-task performance while materially reducing training cost. According to the Reuters report, Alibaba said the model requires roughly one-ninth of the training cost of its previous Qwen3.7 Qwen announced API pricing of 1 yuan (about $0.1488) per million input tokens and 3 yuan per million output tokens for Qwen3.8 Flash access through its application programming interface. Alongside the closed/API Flash release, Qwen also released open-source weights for a sibling variant called Qwen3.8-Flash-Next, which

Artificial Analysis lists Qwen3.8-Flash-Next as an open-weights model from Alibaba, released August 2026, with API provider benchmarking showing Alibaba Cloud as the sole listed host. Measured performance on Alibaba Cloud is 87.5 output tokens per second and a time-to-first-answer-token of 25.44 seconds. The default be Pricing on Alibaba Cloud is reported as a $0.09 blended price per 1M tokens under the default 7:2:1 cache-input-output weighting, which is lower than the per-axis list rates cited elsewhere due to cache-hit weighting methodology. The page separately tracks cache hit, input, and output prices along with a cache discount