实体
model architecture open-source
创建2026-04-06
更新2026-04-10
阅读量级4 分钟
实体导读

Qwen3 is the latest series of open-weight large language models (LLMs) from the Qwen model family, released under Apache 2.0. It includes both dense and Mixture-of-Experts (MoE) architectures with parameter scales ranging from 0.6B to 235B.

  1. Key Innovations
  2. Model Variants
  3. Dense Models
  4. MoE Models
  5. Architecture
  6. Training Pipeline
  7. Pre-training (3 stages)
  8. Post-training (4 stages for flagship)

Qwen3

Qwen3 is the latest series of open-weight large language models (LLMs) from the Qwen model family, released under Apache 2.0. It includes both dense and Mixture-of-Experts (MoE) architectures with parameter scales ranging from 0.6B to 235B.

Key Innovations

  • Thinking Mode and Non-Thinking Mode Integration: A unified framework allowing dynamic switching between complex multi-step reasoning and rapid context-driven responses within a single model, eliminating the need for separate chat-optimized and reasoning models.
  • Thinking Budget: A mechanism enabling users to allocate computational resources adaptively during inference, balancing latency and performance.
  • Strong-to-Weak Distillation: Leveraging knowledge from flagship models to efficiently build smaller-scale models with highly competitive performance.
  • Expanded Multilingual Support: Pre-trained on 36 trillion tokens covering 119 languages and dialects (up from 29 in Qwen2.5).

Model Variants

Dense Models

Model Parameters Context Length
Qwen3-0.6B 0.6B 32K
Qwen3-1.7B 1.7B 32K
Qwen3-4B 4B 128K
Qwen3-8B 8B 128K
Qwen3-14B 14B 128K
Qwen3-32B 32B 128K

MoE Models

Model Total Params Activated Params Context Length
Qwen3-30B-A3B 30B 3B 128K
Qwen3-235B-A22B 235B 22B 128K

Architecture

Based on Qwen2.5 architecture with Grouped Query Attention (GQA), SwiGLU, Rotary Positional Embeddings (RoPE), and RMSNorm with pre-normalization. QKV-bias is removed and QK-Norm is introduced for training stability. MoE models use fine-grained expert segmentation with 128 total experts, 8 activated per token, and no shared experts.

Training Pipeline

Pre-training (3 stages)

  1. General Stage: 30T+ tokens, 4K sequence length, 119 languages
  2. Reasoning Stage: 5T tokens with increased STEM/coding/reasoning data
  3. Long Context Stage: Extended to 32K context using ABF RoPE, YARN, and Dual Chunk Attention

Post-training (4 stages for flagship)

  1. Long-CoT Cold Start
  2. Reasoning RL (GRPO)
  3. Thinking Mode Fusion (SFT)
  4. General-domain RL

Smaller models use Strong-to-Weak Distillation instead of full 4-stage training.

Performance Highlights (Post-trained)

Qwen3-235B-A22B achieves: - AIME'24: 85.7, AIME'25: 81.5 - LiveCodeBench v5: 70.7 - CodeForces: 2056 - BFCL v3: 70.8

Outperforms DeepSeek-V3 Base on 14/15 benchmarks with 1/3 total parameters.

[src: raw/ingested/2025/05/2505.09388.md]