AITOPAI热报

Qwen3

Alibaba Qwen·China·Released 2026-09-30
ClosedGeneral LLM

A conversational large language model developed by Alibaba Qwen, suitable for various dialogue scenarios.

—context
—per 1M tokens (in / out)
—parameters
78stories in our feed

Specs

hover a value to see its source
Context window—
Max output—
Parameters—
Knowledge cutoff—
Licence—
Commercial useUnknown
Input—
Output—
Familyqwen

API price

USD per 1M tokens

No public API price yet.

News timeline

from the AITOP feed, one story per eventAll 78 →
2026-10-21
AWS shows three ways to train Qwen3 on Trainium with torchtitanAWS Machine Learning Blog
2026-10-16
Hojo-ASR-Multi-V1 posts 3.54 average WER across 5 languagesrohanpaul_ai
2026-10-15
Anthropic 论文提出 GRAFT：用可删除适配器替代惩罚来消除 reward hackingrohanpaul_ai
Latent Compaction：用测试时训练压缩 KV cache，32 倍压缩下 RULER 准确率达 96.7%arXiv cs.LG
Behavioral Self-Rewarding Lifts Qwen3-14B on SWE-benchDAIR.AI
2026-10-14
SRO：用LLM智能体自动优化后训练配方，Qwen3上远超基线arXiv: DeepSeek
2026-10-06
研究发现：基础模型加一句前缀即可提升推理成绩Aran Komatsuzaki (论文推介)
2026-10-04
Sebastian Raschka's GRPO Tutorial Lifts Qwen3-0.6B to 47% on MATH-500shao__meng
2026-10-02
Multi-Token Prediction Heads ResearcharXiv: DeepSeek
2026-10-01
微调Qwen3 4B超越Claude Sonnet 4.6Ate-a-Pi
PivotOPD：学习从关键错误中恢复的多轮智能体arXiv cs.AI
2026-09-30
Anthropic指控7家中国公司蒸馏Claude模型掘金本周最热
EASE framework for self-evolving agentsarXiv: DeepSeek
2026-09-29
LLM Post-training Robust to Erroneous RewardsTanishq Abraham (论文推介)
ScopeIF Framework for LLM Instruction FollowingarXiv: DeepSeek
2026-09-28
本地部署大模型何时爆发？作者判断在 2029 年前后Lxfater
研究显示推理模型的思考过程会消解部分偏见但制造更多新偏见arXiv: DeepSeek
2026-09-25
SGLang adds scoring API and multi-item scoring for decision modelsLMSYS Org (SGLang)
2026-09-24
RL-Kernel 集成 vLLM：Qwen3-8B 训练与推理 logprob 逐位一致vLLM
2026-09-21
Open source Kev model released by Cognition teamshao__meng
2026-09-09
LLM-Generated Asset Administration Shell Quality MetricsarXiv: DeepSeek
2026-09-04
ESPO：错误结构化提示优化方法arXiv cs.AI
2026-09-03
UTP-Bench Travel Planning BenchmarkarXiv cs.AI
2026-09-02
Qwen3.8-Max-0902 tops Code Arena阿里通义 Qwen
PersianAnonymizer：评估基于LLM标注的波斯语NER匿名化arXiv: DeepSeek
2026-09-01
Reasoning Models Show Asymmetric DisclosurearXiv: DeepSeek
2026-08-31
Reward-Aligned On-Policy DistillationarXiv: DeepSeek
VISTA：基于验证器的学生-教师自适应蒸馏方法arXiv cs.LG
2026-08-15
用XYZ-Aquila-SFT和Qwen3微调工具调用LLM完整指南marktechpost
2026-08-14
VUI Labs语音模型Luna-TTS登顶TTS Arena，击败ElevenLabsPandaily
MiniMax发布音乐模型MiniMax-Music3，最长可生成5分钟歌曲IT之家
2026-08-11
TIDE：解决LLM在线蒸馏中师生不匹配问题arXiv cs.AI
2026-08-10
TRIAL：智能体强化学习的轨迹相对事后蒸馏框架arXiv cs.LG
2026-08-07
无监督自蒸馏：U-OPSD仅靠模型自身生成提升推理能力arXiv cs.LG
2026-08-06
Skill Entropy：长程推理的技能切换评测与训练arXiv cs.LG
2026-08-05
跨模型KV缓存迁移：线性映射实现预填充复用arXiv cs.LG
2026-08-03
Locus后训练Qwen3超越官方，登顶PostTrainBenchelvis
2026-07-29
Relay-OPD：通过轨迹接力改进同策略蒸馏arXiv cs.AI
MemSFT：用外部参数化内存缓解对齐惩罚arXiv cs.LG
2026-07-24
TTEL：利用错误定位实现推理时扩展arXiv cs.LG