Papersby GlyphOn the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability research paper by Alibaba (Qwen), 2026
Alibaba (Qwen) · Architectures · Aug 31, 2026 · 76 citations · 62 upvotes
What it shows
Qwen3.8-Flash-Next: a 125B mixture-of-experts model with 6B active that nearly matches its 397B predecessor at 1/9 the training compute.
By Zihan Qiu, Zekun Wang, Xiao Li and 33 more · arXiv 2608.30320 · PDF
Summarised by hand from the abstract.
More from Alibaba (Qwen)
All 78Co-Design Bench: Multimodal Collaborative Coding Evaluation for Agentic ModelsCo-Design Bench evaluates how effectively coding agents update software applications based on user requests with text, images, and other media.
Alibaba (Qwen)23 upvotesRead →Qwen-Image-PE: Learning Visual Design for Text-to-Image Prompt EnhancementQwen-Image-PE improves text-to-image prompts via two-stage training with caption-based reverse distillation and groupwise design rewards, scaling across data, model size, and information density.Alibaba (Qwen)23 upvotesRead →World-in-Context: Interactive Video World Modeling with In-Context ConditioningWorld-in-Context unifies view, action, and camera controls via explicit tokens and interaction-aware positional encoding to improve interactive video generation and navigation.Alibaba (Qwen)22 upvotesRead →Beyond Watching and Searching: Benchmarking Agents for Video-centered Deep Research ReportsThis benchmark evaluates how well AI agents can research and synthesize information from videos to produce interactive reports.Alibaba (Qwen)19 upvotesRead →UserArena: Benchmarking Interactive User Simulation for Long-Horizon Shopping AgentsUserArena introduces a real-world interactive benchmark with long-horizon shopping trajectories and an evaluation framework, alongside On-Policy Trajectory Distillation to improve simulator fidelity for agentic recommendations.Alibaba (Qwen)55 upvotesRead →SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolationThis training method uses a smaller, frozen model to guide the training of a larger diffusion model, improving quality without increasing inference costs.Alibaba (Qwen)30 upvotesRead →BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World ReferencesThis benchmark evaluates how well AI models can perform complex 3D creation and editing tasks within the Blender software.Alibaba (Qwen)23 upvotesRead →CommerceGen-Studio: An End-to-End Framework for Commercial-Grade Main-Image GenerationThis paper presents an end-to-end framework for generating commercial-grade e-commerce product images by combining metadata extraction, layout planning, and multi-condition generation.Alibaba (Qwen)25 upvotesRead →
Releases that name it
From releases.fru.devOther architectures papers
TopicAttention Is All You NeedIntroduced the Transformer, the attention-only architecture behind nearly every large language model since.
Google201k citations142 upvotesRead →Mamba: Linear-Time Sequence Modeling with Selective State SpacesA selective state space model that scales linearly with sequence length and matches Transformers on language.
Carnegie Mellon University10.2k citations152 upvotesRead →CodeT5: Identifier-aware Unified Pre-trained Encoder-Decoder Models for Code Understanding and GenerationCodeT5, a unified encoder-decoder Transformer model, improves code understanding and generation by leveraging semantic information from identifiers and user comments.
Salesforce2,633 citations3 upvotesRead →VGGT: Visual Geometry Grounded TransformerVGGT, a feed-forward neural network, efficiently infers multiple 3D attributes from single or multiple views, outperforming alternatives and enhancing downstream tasks without post-processing.
Meta2,132 citations40 upvotesRead →DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language ModelsThe DeepSeekMoE architecture improves expert specialization in Mixture-of-Experts models by segmenting experts and isolating shared ones, achieving better performance and computational efficiency compared to GShard and other models.
DeepSeek1,268 citations65 upvotesRead →CodeT5+: Open Code Large Language Models for Code Understanding and GenerationCodeT5+, a family of flexible encoder-decoder LLMs initialized with frozen LLMs and trained with a mixture of pretraining objectives, achieves state-of-the-art performance across various code-related tasks.Salesforce836 citations7 upvotesRead →
About this paper
- Venue
- arXiv.org
- Citations
- 76, 13 influential · Semantic Scholar
- Upvotes
- 62 · Hugging Face
- Lab
- Alibaba (Qwen) · on Companies · on Quarterly
Changes
Influential citations rose from 11 to 13.
Citations rose from 65 to 76.
Influential citations rose from 5 to 11.
Citations rose from 33 to 65.
Influential citations rose from 2 to 5.
Citations rose from 20 to 33.
Citations rose from 7 to 20.
Citations rose from 6 to 7.
Added to the list.