AITOPAI热报

Qwen3-VL 235B-A22B

Alibaba Qwen·China·Released 2025-04-01
Open weightsDeprecatedMultimodalReasoningTool useVision

Open-weight vision-language model from Alibaba's Qwen team, released April 2025. Accepts text and image input with a 131,072-token context window, supporting chat, reasoning, tool use, and self-hosted deployment.

128Kcontext
$0.7 / $2.8per 1M tokens (in / out)
32.8Bparameters
21stories in our feed

Specs

hover a value to see its source
Context window128K
Max output32K
Parameters32.8B
Knowledge cutoff2025-04
Licenceapache-2.0
Commercial useCommercial OK
InputText · Image
OutputText
HF downloads (30d)2.8M
Familyqwen

API price

USD per 1M tokens
Input$0.7
Output$2.8
Cached input—

No price change recorded since we started tracking.

News timeline

from the AITOP feed, one story per event
2026-10-21
Near-Duplicate 图片选择可向视觉语言模型传递隐含偏好arXiv cs.LG
2026-10-20
OpenVLHarness:带多模态记忆的视觉智能体工具调用框架arXiv cs.AI
WeAct：让智能体通过权重切换获得能力arXiv cs.AI
2026-10-19
vLLM bi-directional KV transfer cuts TTFT in agentic workloadsAWS Machine Learning Blog
2026-10-16
Nick Cammarata essay revisits neuron interpretability using Qwen3-32Bkimmonismus
2026-10-15
Clearcam 开源项目：普通监控摄像头接入 RTSP 即可升级 AI 能力AIGCLINK
AesCode 用图像引导强化学习提升代码生成的视觉质量arXiv cs.AI
EventLens：用事件结构强化学习改进视频时序理解arXiv cs.AI
2026-10-09
WOVEN：视觉世界建模融入多模态大模型arXiv cs.LG
2026-10-06
Infinite-Dreamer：用图像编辑世界模型合成 GUI 智能体训练数据arXiv cs.AI
2026-10-01
BELIEFRAG：自适应RAG在动态证据下的状态感知arXiv cs.AI
2026-09-25
AWS shows SkyRL recipe for post-training Qwen3-VL-8B with GRPOAWS Machine Learning Blog
2026-09-18
Tencent WeChat visual team opens WeVisDoc document parsershao__meng
研究揭示视觉语言模型抗干扰能力与提示词冗长的关系arXiv cs.AI
2026-09-15
Eliciting Native Skill Routing from a Frozen LLMarXiv cs.AI
2026-09-11
Edge-deployable VLMs tested for species identificationarXiv cs.AI
Unified Per-Token Gating Family for On-Policy DistillationarXiv cs.AI
2026-08-05
CARE-X：统一辅助监督与奖励对齐的胸片视觉语言模型arXiv cs.AI
对比激活加和干预实现Qwen3时间偏好双向操控arXiv cs.AI
2026-07-31
本地 Computer-Use Agent 的推理时扩展：失败模式与计算权衡arXiv cs.AI
2026-07-24
视觉对比自蒸馏VCSD：无需外部教师的新方法arXiv cs.AI