Qwen3

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

sparse-MoE full-attention text-only open-weights frontier frontier-generalist

Design overview

Qwen3's MoE branch pairs a deliberately conventional transformer backbone with an aggressive sparsity story. Qwen3-235B-A22B and Qwen3-30B-A3B share the dense branch's Qwen2.5-style stack — GQA attention with QK-Norm, SwiGLU, RoPE with the base raised to 1,000,000 via ABF, pre-RMSNorm — but replace every FFN with a fine-grained MoE: 128 total experts with 8 activated per token and, unlike Qwen2.5-MoE or DeepSeek-V3, no shared expert at all. The report discloses no hidden or per-expert dimensions, no router function, and no embedding tying, so the block's economics are only partially visible; what is stated is that load is balanced by a global-batch load-balancing auxiliary loss (Qiu et al., 2025). Context reaches 128K via YaRN + Dual Chunk Attention over the 32,768-token stage-3 training length. The 235B flagship activates 22B parameters per token — the same active count as DeepSeek-V3 at roughly a third of its total parameters — and the family's headline claims sit in post-training: one model that switches between thinking and non-thinking modes via /think and /no_think chat-template flags, a thinking-budget mechanism that halts reasoning at a user-defined threshold, and strong-to-weak distillation that builds the lightweight members (including the 30B-A3B MoE) at roughly 1/10 of the four-stage RL GPU hours. The flagship itself runs the full pipeline: long-CoT cold start, GRPO reasoning RL on 3,995 query-verifier pairs, thinking-mode fusion SFT, then general RL with rule-based and model-based rewards.

What the report claims is novel

Atlas assessment: Architecturally the MoE branch is conservative — a Qwen2.5-lineage GQA + SwiGLU + RoPE + pre-RMSNorm transformer (plus QK-Norm) with fine-grained routing. Against DeepSeek-V3, the closest atlas peer (671B total / 37B active, 256 routed experts with 1 shared expert and aux-loss-free bias balancing), Qwen3-MoE is the mirror-image design: fewer and fatter experts (128 total / 8 active), no shared expert at all, and an explicit global-batch load-balancing auxiliary loss rather than an aux-loss-free scheme. Both choices (fine-grained segmentation, global-batch loss) are borrowed and cited (Dai et al. 2024; Qiu et al. 2025), not new mechanisms. The genuinely distinctive contributions sit in post-training — thinking/non-thinking mode unification, the thinking budget, and strong-to-weak distillation — and in scale economics: 22B active parameters matching DeepSeek-V3's active count at 35% of its total parameters. The report's efficiency claims (1/5 activated params vs dense, 1/10 GPU hours for distillation) are internal comparisons, not independently verifiable.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

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 parameters235B[7]
Active parameters / token22B[8]
Layers94[9]
Hidden dimensionnot disclosed
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query heads64[10]
KV heads4[10]
Head dimensionnot disclosed
Vocabulary size151669[11]
Embedding tyingnot disclosed

Per-variant configs

VariantTotalLayersHiddenFFNQ headsKV headsContext
Qwen3-30B-A3B30B48not disclosednot disclosed324131,072
Qwen3-235B-A22B235B94not disclosednot disclosed644131,072

Core block

Block typesparse-MoE
Attention variantGQA[12]
Attention layer patternuniform
Depth mixingsequential-residual
ActivationSwiGLU[13]
Expert count128[14]
Experts per token8[15]
Shared experts0[16]
Routingnot disclosed
Load balancingglobal-batch load balancing loss (Qiu et al., 2025) to encourage expert specialization[13]
Expert granularityfine-grained expert segmentation following Qwen2.5-MoE (Dai et al., 2024)[17]
MethodRoPE[20]
RoPE base frequency1000000[21]
Partial RoPEnot disclosed
TypeRMSNorm[22]
Placementpre[23]
QK-normyes[24]
Attention sinksnot disclosed
Softcappingnot disclosed
Othernot disclosed

Context

Trained context length32,768[25]
Deployed context length131,072
Extension methodYaRN + Dual Chunk Attention (DCA) with ABF base-frequency increase to 1,000,000; four-fold inference length extension[26]

Tokenizer

Algorithmbyte-level BPE (BBPE)[27]
NotesQwen tokenizer, vocabulary 151,669; chat template adds /think and /no_think flags and <think>/</think> tokens for thinking-mode control (Table 9)[28]

Training

Training tokens36T[29]
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)[30]
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)[26]
Optimizernot disclosed
LR schedulelearning-rate decay accelerated during S2; optimal learning rate per model predicted via scaling laws (values not printed)[26]
Batch schedulebatch size strategy predicted via scaling laws for each dense/MoE model (values not printed)[26]
Precisionnot disclosed
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[31]
Preference optimisationGRPO[32]
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[33]
DistillationQwen3-30B-A3B: strong-to-weak distillation (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). Flagship Qwen3-235B-A22B is not distilled — it runs the full four-stage post-training pipeline.[34]

Modality

Typetext-only
Attachmentn/a[35]

Inference efficiency

KV-cache designGQA: KV cache scales with 4 KV heads (Qwen3-235B-A22B: 64 Q / 4 KV heads; Qwen3-30B-A3B: 32 Q / 4 KV heads)
Quantisation shippednone disclosed[36]
Speculative decodingnone disclosed[36]
Serving optimisationsnone disclosed[37]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
MMLU (5-shot)87.81Table 3
MMLU-Pro (5-shot, CoT)68.18Table 3
BBH (3-shot, CoT)88.87Table 3
GPQA (5-shot, CoT)47.47Table 3
MATH (4-shot, CoT)71.84Table 3
EvalPlus (0-shot)77.6Table 3
MBPP (3-shot)81.4Table 3
MMLU (5-shot)81.38Table 5
MMLU-Pro (5-shot, CoT)61.49Table 5
MATH (4-shot, CoT)59.04Table 5
EvalPlus (0-shot)71.45Table 5
AIME'24 (Thinking)85.7Table 11
AIME'25 (Thinking)81.5Table 11
GPQA-Diamond (Thinking)71.1Table 11
Arena-Hard (Thinking)95.6Table 11
LiveCodeBench v5 (Thinking)70.7Table 11
BFCL v3 (Thinking)70.8Table 11
CodeForces (Rating / Percentile, Thinking)2056 / 98.2%Table 11
MMLU-Redux (Non-thinking)89.2Table 12
Arena-Hard (Non-thinking)96.1Table 12
MATH-500 (Non-thinking)91.2Table 12
LiveCodeBench v5 (Non-thinking)35.3Table 12
AIME'24 (Thinking)80.4Table 15
MATH-500 (Thinking)98.0Table 15
ZebraLogic (Thinking)89.5Table 15
RULER (Avg., Non-thinking)95.0Table 23
RULER (Avg., Non-thinking)91.6Table 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 · §1, §2 (flagship 'has a total of 235B parameters'); Table 3 (# Total Params 235B)
  8. stated · §1 ('22 billion activated ones per token'); Table 3 (# Activated Params 22B)
  9. stated · Table 2 (Qwen3-235B-A22B)
  10. stated · Table 2 (Heads Q/KV: 64 / 4 for Qwen3-235B-A22B)
  11. stated · §2 ('a vocabulary size of 151,669')
  12. stated · §2 (MoE models 'share the same fundamental architecture as the Qwen3 dense models', which use Grouped Query Attention; Table 2 heads Q/KV 64/4 and 32/4)
  13. stated · §2
  14. stated · §2 ('128 total experts'); Table 2 (# Experts Total / Activated: 128 / 8)
  15. stated · §2 ('8 activated experts per token'); Table 2 (# Experts Total / Activated: 128 / 8)
  16. stated · §2 ('Unlike Qwen2.5-MoE, the Qwen3-MoE design excludes shared experts')
  17. stated · §2 ('We follow Qwen2.5-MoE and implement fine-grained expert segmentation (Dai et al., 2024)')
  18. stated · §3.2, §A.1.1
  19. stated · §A.1.1 ('YARN with a scaling factor=4'); §3.2 ('a four-fold increase in sequence length capacity during inference')
  20. stated · §2 (Rotary Positional Embeddings)
  21. stated · §3.2 ('we increase the base frequency of RoPE from 10,000 to 1,000,000 using the ABF technique')
  22. stated · §2 (RMSNorm with pre-normalization)
  23. stated · §2 ('RMSNorm with pre-normalization')
  24. stated · §2 (QK-Norm (Dehghani et al., 2023) introduced to the attention mechanism; 'The Qwen3 MoE models share the same fundamental architecture as the Qwen3 dense models')
  25. stated · §3.2 stage 3 ('All models are pre-trained on hundreds of billions of tokens with a sequence length of 32,768 tokens')
  26. stated · §3.2
  27. stated · §2 ('Qwen's tokenizer ... implements byte-level byte-pair encoding (BBPE)')
  28. stated · §2, §4.3
  29. stated · §1, §3.1 ('a total of 36 trillion tokens' for the whole Qwen3 family)
  30. stated · §3.1
  31. stated · §4.1 (long-CoT cold-start SFT), §4.3 (continual SFT in Thinking Mode Fusion)
  32. 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 (§4.4)
  33. stated · §4.1–4.3
  34. stated · §4.5, §4.7 (Table 21)
  35. n/a · §2 (text-only architecture)
  36. stated · report-wide absence
  37. stated · report-wide absence (YaRN deployment to 64K context appears only in BFCL evaluation settings, §4.6)