Qwen3

Qwen Team · 2025-05-15 · analysed 2026-08-10 · skill v0.3.5

dense full-attention text-only open-weights large frontier-generalist

Design overview

Qwen3 dense is the standard-bearer of the Qwen2.5 architectural line: each layer is a pre-RMSNorm transformer block with GQA attention, SwiGLU feed-forward, and RoPE, stacked 64 deep in the reference Qwen3-32B (36 layers in 4B/8B, 40 in 14B, 28 in 0.6B/1.7B). The two structural changes over Qwen2.5 are the removal of QKV-bias and the addition of QK-Norm to the attention mechanism, both aimed at training stability. Context is grown in three pre-training stages (4,096 → 4,096 → 32,768 tokens, 36T tokens total across 119 languages), then extended four-fold at inference via YaRN with scaling factor 4 plus Dual Chunk Attention, on top of an ABF base-frequency increase from 10,000 to 1,000,000; the 32B deploys at 128K. Post-training is where Qwen3 makes its mark: a long-CoT cold start with QwQ-32B-generated reasoning traces, GRPO-based reasoning RL on 3,995 verifiable query-verifier pairs, a thinking-mode fusion stage that fuses /think and /no_think behaviors into one model with a budget-controllable thinking process, and a general RL stage with rule-based and model-based rewards. The five smaller dense siblings (0.6B–14B) skip most of this: they are trained via strong-to-weak distillation, first off-policy on teacher outputs and then on-policy against teacher logits, at roughly a tenth of the RL compute.

What the report claims is novel

Atlas assessment: The dense block is a conservative Qwen2.5-lineage stack (GQA + SwiGLU + RoPE + RMSNorm pre-norm) with one structural addition, QK-Norm, and one removal, QKV-bias; closest atlas relative is meta-llama-3.1 (same four-mechanism recipe, also RoPE + pre-RMSNorm), from which Qwen3 dense differs mainly by QK-Norm, the 1,000,000 RoPE base with YaRN+DCA 4× extension instead of direct 128K training, and a tie-embedding policy split by size (small models tie, 8B+ do not). The genuinely novel contribution is the post-training pipeline (long-CoT cold start → reasoning RL with GRPO → thinking-mode fusion → general RL) and strong-to-weak distillation for small models; these are training recipes, not new mixing mechanisms. The report discloses no hidden dims or FFN dims (unlike Llama 3.1 / Phi-3 / DeepSeek-V3 reports), so per-variant width comparisons in the atlas are impossible from this source.

Relation to the atlas

Lineage in

Influence out

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyQwen3[1]
OrganisationQwen Team[2]
Release date2025-05-15[3]
Report URLhttps://arxiv.org/abs/2505.09388[4]
LicenseApache 2.0[5]
Open weightsyes[6]

Variants

Scale

Total parameters32B[7]
Active parameters / token32B
Layers64[8]
Hidden dimensionnot disclosed
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query heads64[9]
KV heads8[9]
Head dimensionnot disclosed
Vocabulary size151669[10]
Embedding tyingno[11]

Per-variant configs

VariantTotalLayersHiddenFFNQ headsKV headsContext
Qwen3-0.6B600M28not disclosednot disclosed16832,768
Qwen3-1.7B1.7B28not disclosednot disclosed16832,768
Qwen3-4B4B36not disclosednot disclosed328131,072
Qwen3-8B8B36not disclosednot disclosed328131,072
Qwen3-14B14B40not disclosednot disclosed408131,072

Core block

Block typedense
Attention variantGQA[12]
Attention layer patternuniform
Depth mixingsequential-residual
ActivationSwiGLU[13]
Expert countn/a[14]
Experts per tokenn/a[14]
Shared expertsn/a[14]
Routingn/a[14]
Load balancingn/a[14]
Expert granularityn/a[14]
MethodRoPE[17]
RoPE base frequency1000000[18]
Partial RoPEnot disclosed
TypeRMSNorm[19]
Placementpre[20]
QK-normyes[21]
Attention sinksnot disclosed
Softcappingnot disclosed
Othernot disclosed

Context

Trained context length32,768[22]
Deployed context length131,072
Extension methodYaRN + Dual Chunk Attention (DCA) with ABF base-frequency increase to 1,000,000; four-fold inference length extension[23]

Tokenizer

Algorithmbyte-level BPE (BBPE)[24]
NotesQwen tokenizer, vocabulary 151,669; chat template adds /think and /no_think flags and <think>/</think> tokens for thinking-mode control (Table 9)[25]

Training

Training tokens36T[26]
Data composition119 languages and dialects; domains include coding, STEM, reasoning tasks, books, multilingual texts, and synthetic data; PDF-like documents transcribed by Qwen2.5-VL with Qwen2.5 refinement; trillions of synthetic tokens generated by Qwen2.5, Qwen2.5-Math, Qwen2.5-Coder (textbooks, QA, instructions, code); instance-level data-mixture optimization via a multilingual annotation system applied to over 30 trillion tokens (educational value, fields, domains, safety)[27]
Curriculum / stagingthree-stage pre-training: S1 general stage >30T tokens at sequence length 4,096; S2 reasoning stage ~5T higher-quality tokens at 4,096 with accelerated learning-rate decay; S3 long-context stage hundreds of billions of tokens at 32,768 (75% 16,384–32,768-token texts, 25% 4,096–16,384)[23]
Optimizernot disclosed
LR schedulelearning-rate decay accelerated during S2; optimal learning rate per model predicted via scaling laws (values not printed)[23]
Batch schedulebatch size strategy predicted via scaling laws for each dense/MoE model (values not printed)[23]
Precisionnot disclosed
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[28]
Preference optimisationGRPO[29]
Reasoning traininglong-CoT cold start (responses generated by QwQ-32B, two-phase query/response filtering); Reasoning RL on 3,995 query-verifier pairs (math/code/STEM); thinking budget mechanism halting thinking at a user-defined threshold; /think and /no_think mode switching with <think>/</think> tokens[30]
DistillationStrong-to-Weak Distillation for smaller models (0.6B/1.7B/4B/8B/14B dense and 30B-A3B MoE): off-policy response distillation from teacher outputs, then on-policy distillation aligning student logits with teacher (Qwen3-32B or Qwen3-235B-A22B) via KL divergence; ~1/10 the GPU hours of four-stage RL[31]

Modality

Typetext-only
Attachmentn/a[32]

Inference efficiency

KV-cache designGQA: KV cache scales with 8 KV heads (Qwen3-32B: 64 Q heads / 8 KV heads)
Quantisation shippednone disclosed[33]
Speculative decodingnone disclosed[33]
Serving optimisationsnone disclosed[34]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
MMLU (5-shot)83.61Table 4
MMLU-Pro (5-shot, CoT)65.54Table 4
SuperGPQA (5-shot, CoT)39.78Table 4
BBH (3-shot, CoT)87.38Table 4
GPQA (5-shot, CoT)49.49Table 4
GSM8K (4-shot, CoT)93.4Table 4
MATH (4-shot, CoT)61.62Table 4
EvalPlus (0-shot)72.05Table 4
MultiPL-E (0-shot)67.06Table 4
MBPP (3-shot)78.2Table 4
CRUX-O (1-shot)72.5Table 4
MGSM (8-shot, CoT)83.06Table 4
MMMLU (5-shot)83.83Table 4
INCLUDE (5-shot)67.87Table 4
MMLU-Redux (Thinking)90.9Table 13
GPQA-Diamond (Thinking)68.4Table 13
AIME'24 (Thinking)81.4Table 13
AIME'25 (Thinking)72.9Table 13
LiveBench 2024-11-25 (Thinking)74.9Table 13
Arena-Hard (Thinking)93.8Table 13
BFCL v3 (Thinking)70.3Table 13
LiveCodeBench v5 (Thinking)65.7Table 13
CodeForces Rating (Thinking)1977Table 13
RULER 128K (Non-thinking)85.6Table 23

Provenance references

  1. stated · Title
  2. stated · Title page byline (huggingface.co/Qwen, github.com/QwenLM/Qwen3)
  3. stated · Title page date line (2025-05-15); arXiv stamp reads 14 May 2025
  4. stated · pipeline
  5. stated · Abstract ('all Qwen3 models are publicly accessible under Apache 2.0')
  6. stated · Abstract, §1 ('open-weight large language models'; 'publicly accessible under Apache 2.0')
  7. stated · Table 4 (# Total Params 32B); model name Qwen3-32B
  8. stated · Table 1 (Qwen3-32B)
  9. stated · Table 1 (Heads Q/KV: 64 / 8 for Qwen3-32B)
  10. stated · §2 ('a vocabulary size of 151,669')
  11. stated · Table 1 (Tie Embedding: No for Qwen3-32B; Yes only for 0.6B/1.7B/4B)
  12. stated · §2 (Grouped Query Attention, GQA)
  13. stated · §2
  14. n/a · §2 (dense models; no MoE)
  15. stated · §3.2, §A.1.1
  16. stated · §A.1.1 ('YARN with a scaling factor=4'); §3.2 ('a four-fold increase in sequence length capacity during inference')
  17. stated · §2 (Rotary Positional Embeddings)
  18. stated · §3.2 ('we increase the base frequency of RoPE from 10,000 to 1,000,000 using the ABF technique')
  19. stated · §2 (RMSNorm with pre-normalization)
  20. stated · §2 ('RMSNorm with pre-normalization')
  21. stated · §2 ('we remove QKV-bias used in Qwen2 and introduce QK-Norm (Dehghani et al., 2023) to the attention mechanism to ensure stable training')
  22. stated · §3.2 stage 3 ('All models are pre-trained on hundreds of billions of tokens with a sequence length of 32,768 tokens')
  23. stated · §3.2
  24. stated · §2 ('Qwen's tokenizer ... implements byte-level byte-pair encoding (BBPE)')
  25. stated · §2, §4.3
  26. stated · §1, §3.1 ('a total of 36 trillion tokens')
  27. stated · §3.1
  28. stated · §4.1 (long-CoT cold-start SFT), §4.3 (continual SFT in Thinking Mode Fusion)
  29. stated · §4.2 ('employed GRPO (Shao et al., 2024) to update the model parameters'); Stage 4 General RL adds rule-based and model-based rewards
  30. stated · §4.1–4.3
  31. stated · §4.5, §4.7 (Table 21)
  32. n/a · §2 (text-only architecture)
  33. stated · report-wide absence
  34. stated · report-wide absence (YaRN deployment to 64K context appears only in BFCL evaluation settings, §4.6)