GLM-4.5 (ARC series)

Zhipu AI & Tsinghua University (GLM-4.5 Team) · 2025-08-08 · analysed 2026-08-10 · skill v0.3.5

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

Design overview

GLM-4.5-Air is the compact sibling in Zhipu AI's GLM-4.5 series: a 46-layer sparse MoE (45 MoE + 1 dense layer) with 106B total and 12B activated parameters, hidden dim 4096, 128 routed experts (8 active) plus one shared expert, and per-expert intermediate dim 1408 against 10944 in its single dense layer. It shares the flagship's block design — GQA with 96 query heads over 8 KV heads (head dim 128), partial RoPE with the base raised from 10,000 to 1,000,000 at the 32K context stage, sigmoid gating with loss-free bias-based load balancing plus a small sequence-level balance loss, RMSNorm, and a dedicated MoE MTP layer used for speculative decoding — with one notable exception: Table 1 marks QK-Norm as 'No' for Air, the only stated architectural difference from GLM-4.5. Training follows the same series recipe (two-stage pre-training, repo-level code / synthetic reasoning / long-context-agent mid-training, 4K→128K context growth, Muon optimizer, cosine decay, BF16 with FP8 rollouts), though the report gives no token budget specific to Air. Post-training is the shared expert-model-iteration pipeline: domain experts for Reasoning, Agent, and General chat are trained and then distilled via SFT into one hybrid-reasoning model with thinking and non-thinking modes. The report positions Air as a 100B-scale parameter-efficiency play, matching or exceeding Qwen3-235B-A22B and MiniMax-M1 on ARC benchmarks while activating only 12B parameters.

What the report claims is novel

Atlas assessment: GLM-4.5-Air is the same block as GLM-4.5 at smaller scale (4096 hidden, 46 layers, 128 experts) with exactly one stated architectural difference: QK-Norm is absent (Table 1). The record exists separately from zai-glm-4-5 because its config differs, not because its block type differs — the task premise that Air is 'dense' is contradicted by the report, which explicitly calls it an MoE ('GLM-4.5-Air is a smaller MoE model with 106B parameters', §1; 45/46 MoE layers, Table 1). Relative to the 100B-scale atlas class (qwen-qwen3-moe 235B-A22B), Air is narrower and deeper with 96/8 GQA heads and per-expert width 1408; its 12B active params are well below Qwen3's 22B, which the report leverages as a parameter-efficiency claim. The QK-Norm ablation between flagship and Air is disclosed but never analyzed, which weakens the causal story for Air's lower scores.

Relation to the atlas

Lineage in

Influence out

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyGLM-4.5 (ARC series)[1]
OrganisationZhipu AI & Tsinghua University (GLM-4.5 Team)[2]
Release date2025-08-08
Report URLhttps://arxiv.org/abs/2508.06471[3]
Licensenot disclosed
Open weightsyes[4]

Variants

Scale

Total parameters106B[5]
Active parameters / token12B[5]
Layers46
Hidden dimension4096[6]
FFN inner dimension1408[7]
FFN ratio0.34
Query heads96[8]
KV heads8[9]
Head dimension128[10]
Vocabulary sizenot disclosed
Embedding tyingnot disclosed

Per-variant configs

VariantTotalLayersHiddenFFNQ headsKV headsContext
GLM-4.5355B9251201536968131,072

Core block

Block typesparse-MoE[11]
Attention variantGQA[12]
Attention layer patternuniform[13]
Depth mixingsequential-residual
Activationnot disclosed
Expert count128[14]
Experts per token8[15]
Shared experts1[16]
Routingsigmoid gates with loss-free balance routing[17]
Load balancingloss-free balance routing: expert bias update rate 0.001 for the first 15T tokens, then 0.0; auxiliary sequence-level balance loss with weight 0.0001[18]
Expert granularitycompact per-expert FFN (MoE intermediate dim 1408, 0.34× hidden) vs dense-layer intermediate 10944; report uses no fine-grained/grouped terminology
MethodRoPE[22]
RoPE base frequency1000000[23]
Partial RoPEyes[24]
TypeRMSNorm[25]
Placementnot disclosed
QK-normno[26]
Attention sinksno
Softcappingno

Context

Trained context length131,072[27]
Deployed context length131,072[28]
Extension methodtrained at extended lengths: 4K → 32K → 128K across pre-/mid-training, with RoPE base frequency raised from 10,000 to 1,000,000 at the 32K stage[19]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokensnot disclosed
Data compositionseries-level corpus (report describes the GLM-4.5 series jointly): webpages, social media, books, papers, and code repositories; web documents bucketed by quality (top bucket >3.2 epochs, lowest discarded, MinHash + SemDedup); multilingual from crawled web and Fineweb-2 with quality-classifier up-sampling; code from GitHub with tiered quality models and Fill-In-the-Middle; math & science up-sampled above an LLM-scored threshold; two-stage pre-training (general documents, then up-sampled code/math/science)[29]
Curriculum / stagingtwo-stage pre-training; mid-training stages: repo-level code (4K→32K), synthetic reasoning data, long-context & agent training (32K→128K); best-fit packing in mid-training only (series-level recipe)[30]
OptimizerMuon for all parameters except word embeddings, biases, and RMSNorm weights (Newton-Schulz iterations N=5, momentum 0.95, scaled update RMS 0.2); weight decay 0.1; no dropout (series-level recipe)[31]
LR schedulecosine decay (chosen over warmup-stable-decay); warmup 0 → 2.5e-4, decaying to 2.5e-5 until the end of mid-training (series-level recipe)[31]
Batch schedulebatch size warmup: 16M → 64M tokens over the first 500B tokens, then constant (series-level recipe)[31]
PrecisionBF16 training with FP8 inference for RL rollouts (online block-wise FP8 quantization before rollout dispatch)[32]
ParallelismMegatron-based training (per Slime infra description); detailed TP/PP/EP/CP strategy not described[33]
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[34]
Preference optimisationGRPO[35]
Reasoning traininghybrid reasoning modes (thinking + non-thinking); reasoning RL with two-stage difficulty-based curriculum, single-stage RL at 64K max output length, dynamic sampling temperature, token-weighted mean loss for code RL; agentic RL (web-search + SWE) with process format penalty and iterative self-distillation; general RL (holistic, instruction-following, function-calling, pathology); test-time compute via interaction turns (series-level pipeline, §3.2–3.4)[36]
Distillationself-distillation (no external teacher): Stage 1 domain experts (Reasoning, Agent, General chat) consolidated into a unified hybrid-reasoning model in Stage 2 via SFT distillation; iterative self-distillation in agentic RL (series-level pipeline)[37]

Modality

Typetext-only
Attachmentn/a[38]

Inference efficiency

KV-cache designGQA with 8 KV heads shared across 96 query heads: KV cache proportional to 8 heads (12× reduction vs 96-head MHA)
Quantisation shippednone disclosed[32]
Speculative decodingMTP (Multi-Token Prediction) MoE layer supporting speculative decoding during inference; MTP loss weight λ = 0.3 for the first 15T tokens, 0.1 after (series-level recipe)[39]
Serving optimisationsnone disclosed[32]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
TAU-Bench (Retail / Airline)77.9 / 60.8Table 3
BFCL V376.4Table 3
BrowseComp21.3Table 3
AIME 24 (Avg@32)89.4Table 4
GPQA (Avg@8)75.0Table 4
MMLU-Pro81.4Table 4
MATH 50098.1Table 4
SciCode37.3Table 4
HLE10.6Table 4
LiveCodeBench (2407-2501)70.7Table 4
SWE-bench Verified57.6Table 5
Terminal-Bench30.0Table 5
MMLU87.4Table 6
SimpleQA14.5Table 6
IFEval (Prompt Strict)86.3Table 6
SysBench (ISR)77.4Table 6
MultiChallenge42.5Table 6
SafetyBench (average)87.8Table 7
Novel logical reasoning (expert score)53.4Table 11

Provenance references

  1. stated · Title; §1 ('This paper introduces two new models: GLM-4.5 and GLM-4.5-Air')
  2. stated · Title page byline
  3. stated · pipeline
  4. stated · Abstract ('We release both GLM-4.5 (355B parameters) and a compact version, GLM-4.5-Air (106B parameters)'); §1 (huggingface.co/zai-org/GLM-4.5); §5 ('We release the model weights of GLM-4.5 and GLM-4.5-Air')
  5. stated · Abstract; Table 1
  6. stated · Table 1 (Hidden Dim)
  7. stated · Table 1 (MoE Intermediate Dim)
  8. stated · Table 1 (# Attention Heads)
  9. stated · Table 1 (# Key-Value Heads) — GQA
  10. stated · Table 1 (Attention Head Dim)
  11. stated · §1 ('GLM-4.5-Air is a smaller MoE model with 106B parameters'); §5 ('Both models adopt the MoE architecture'); Table 1
  12. stated · §2.1 ('Grouped-Query Attention with partial RoPE'); Table 1 (96 Q heads / 8 KV heads)
  13. stated · §2.1 (GQA in the self-attention component; no per-layer hybrid described)
  14. stated · Table 1 (# Experts (total) 128)
  15. stated · Table 1 (# Experts Active Per Token)
  16. stated · Table 1 (# Shared Experts)
  17. stated · §2.1 ('In the GLM-4.5 series, we adopt the MoE architecture... We employ loss-free balance routing [40] and sigmoid gates for MoE layers [23]')
  18. stated · §2.1, §2.4 (series-level recipe)
  19. stated · §2.3, §2.4
  20. n/a · §2.4 (base-frequency change, not an interpolation/rescaling factor)
  21. stated · §2.4 (max sequence length extended to 32,768 and 131,072 during mid-training)
  22. stated · §2.1 (partial RoPE); §2.4 (RoPE base frequency adjustment)
  23. stated · §2.4 ('we also adjusted RoPE's base frequency from 10,000 to 1,000,000' when extending to 32K)
  24. stated · §2.1 (Grouped-Query Attention with partial RoPE)
  25. stated · §2.4 ('weights for RMSNorm' excluded from the Muon optimizer; series-level recipe)
  26. stated · Table 1 (QK-Norm: No)
  27. stated · §2.4 (series-level recipe: max sequence length 4,096 during pre-training, extended to 32,768 and 131,072 in mid-training); §3.1 (overall SFT at max 128K)
  28. stated · §4.2.3 (evaluation 'prevent exceeding the 128K context limit')
  29. stated · §2.2
  30. stated · §2.2, §2.3
  31. stated · §2.4
  32. stated · §3.5
  33. stated · §3.5 (Fig. 10)
  34. stated · §3.1 (cold-start SFT + overall SFT at 128K max context; series-level pipeline applying to both models)
  35. stated · §3.2 ('Our overall RL algorithm builds upon the GRPO [31] framework, excluding the KL loss term'); §3.4
  36. stated · §1, §3.2, §3.3, §3.4
  37. stated · §3, §3.3.2
  38. n/a · §1–§4 (text-only model)
  39. stated · §2.1, §2.4