GLM-5

Zhipu AI & Tsinghua University (GLM-5 Team) · 2026-02-17 · analysed 2026-08-10 · skill v0.3.5

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

Design overview

GLM-5 is a 78-layer (3 dense + 75 MoE), 744B-parameter sparse MoE with 40B active parameters and 8 of 256 experts (1 shared) active per token. Its defining move is replacing the dense attention of GLM-4.5 with DeepSeek Sparse Attention on every layer: a lightning indexer (32 heads, head dim 128) scores all cached KV entries per query and attention is computed only over the top-k = 2048 retrieved tokens, cutting long-sequence attention cost ~1.5-2×. The attention itself remains DeepSeek-style MLA, re-tuned as MLA-256 (64 heads, QK head dim 192, V head dim 256, 576-dim latent KV) with an added Q LoRA, and with Muon Split — per-head orthogonalization of the attention up-projections — stabilizing logit scale without clipping. Decoding is sped by sharing parameters across 3 MTP layers, which also serve as the speculative draft. Context grows 4K → 200K through a staged mid-training curriculum (32K/128K/200K) with long-context agentic data, plus a DSA warm-up and 20B-token sparse-adaptation stage; SFT then runs at 202,752-token windows. Post-training is a sequential Reasoning → Agentic → General RL pipeline (GRPO-based with IcePop-style mismatch control; asynchronous, decoupled rollout engines with a TITO gateway and direct double-sided importance sampling for agentic work) closed by on-policy cross-stage distillation. The report's framing is 'vibe coding to agentic engineering': long-horizon SWE, terminal, search and slide-generation environments, CC-Bench-V2 evaluation, and a full-stack adaptation to seven Chinese GPU platforms with W4A8 quantization. Note: the gallery's GLM-5.1 and GLM-5.2 cards cite this report, but the report itself discloses only GLM-5.

What the report claims is novel

Atlas assessment: The block is a conservative evolution of the GLM-4.5/DeepSeek-V3 MLA-MoE lineage: the same 256-total/8-active/1-shared expert organization as DeepSeek-V3, MLA retained but re-tuned (MLA-256 with a Q LoRA added, Muon Split for head-wise optimizer scaling), and the dense-attention backbone replaced by DeepSeek-V3.2's DSA sparse attention on every layer. Against atlas peers: vs deepseek-v3 the differences are DSA, Muon (not AdamW), MTP parameter sharing, and the 3-dense-layer structure; vs zai-glm-4-5 it roughly doubles parameters (355B→744B) while shrinking depth (92→78 layers) and swapping in DSA. None of the attention-side ingredients is original to this report (MLA and DSA are both credited to DeepSeek), and the sparse-attention RL-stability fix (deterministic top-k, frozen indexer) is a pragmatic engineering contribution rather than a new mechanism. The genuinely distinctive contributions are in the post-training and serving systems: fully asynchronous agentic RL with TITO gateways and direct double-sided importance sampling, on-policy cross-stage distillation, and the multi-vendor Chinese-GPU inference stack. Reported gains ('~20% over GLM-4.7', '50 on AA Index v4.0') are internal or third-party leaderboard claims and not independently verifiable from the report.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyGLM-5[1]
OrganisationZhipu AI & Tsinghua University (GLM-5 Team)[2]
Release date2026-02-17
Report URLhttps://arxiv.org/abs/2602.15763[3]
Licensenot disclosed
Open weightsyes[4]

Scale

Total parameters744B[5]
Active parameters / token40B[6]
Layers78
Hidden dimension6144[7]
FFN inner dimension2048[8]
FFN ratio0.33
Query heads64[9]
KV headsn/a[10]
Head dimension192[11]
Vocabulary size154880[12]
Embedding tyingnot disclosed

Core block

Block typesparse-MoE
Attention variantMLA[13]
Attention layer patternuniform — DeepSeek Sparse Attention (DSA) applied to all layers: a lightning indexer (32 heads, head dim 128, top-k = 2048 KV entries per query) selects relevant keys/values and attention is computed sparsely over the retrieved subset; no dense-attention or sliding-window alternation[14]
Depth mixingsequential-residual
Activationnot disclosed
Expert count255
Experts per token8[15]
Shared experts1[16]
Routingnot disclosed
Load balancingnot disclosed
Expert granularitynot disclosed
Methodnot disclosed
RoPE base frequencynot disclosed
Partial RoPEnot disclosed
Typenot disclosed
Placementnot disclosed
QK-normno
Attention sinksno
Softcappingno

Context

Trained context length200K[19]
Deployed context length200K[20]
Extension methodprogressive mid-training curriculum (32K → 128K → 200K) plus DSA sparse adaptation; no positional-encoding rescaling disclosed[17]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens28.5T[21]
Data compositionWeb: refined GLM-4.5 pipeline with an additional DCLM sentence-embedding classifier and a World Knowledge classifier (Wikipedia/LLM-labeled) for long-tail knowledge. Code: refreshed snapshots from code hosts plus code-containing web pages (+28% fuzzily deduplicated unique tokens), Software Heritage metadata fixes, and dedicated classifiers for low-resource programming languages (Scala, Swift, Lua, etc.). Math & science: LLM-scored documents from webpages/books/papers with chunk-and-aggregate scoring for long documents; synthetic/AI-generated/template data strictly filtered out[22]
Curriculum / stagingtwo-stage base training (pre-training prioritizing code and reasoning early on, then mid-training for agentic and long-context capacity) with progressive context extension 32K (1T tokens) → 128K (500B) → 200K (50B); long-context data (natural + synthetic interleaved packing inspired by NextLong/EntropyLong, MRCR-like data at 200K) upsampled at later stages; DSA warm-up (1,000 steps, 14 sequences of 202,752 tokens, lr 5e-3 → 2e-4) then sparse adaptation on 20B tokens (constant lr 1e-5)[23]
OptimizerMuon with Muon Split (per-head Newton-Schulz-style matrix orthogonalization applied to independent per-head up-projection matrices); zero-redundant distributed communication variant[24]
LR schedulepre-training: warmup from 0 to 2e-4 then decay to 4e-5 (cosine per GLM-4.5 setting); mid-training: linear decrease 4e-5 → 1e-5; DSA warm-up: 5e-3 → 2e-4; DSA sparse adaptation: constant 1e-5[25]
Batch schedulenot disclosed
Precisionnot disclosed
Parallelisminterleaved pipeline parallelism with flexible MTP placement (MTP output layer co-located with main output layer for parameter sharing), Pipeline ZeRO2 gradient sharding with double-buffered accumulation buffers, zero-redundant all-gather for the Muon optimizer, layer-granular pipeline activation offloading to host memory, sequence-chunked output projection, workload-aware sequence reordering with dynamic redistribution and flexible context-parallel groups using hierarchical all-to-all[26]
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[27]
Preference optimisationGRPO[28]
Reasoning trainingsequential RL pipeline: Reasoning RL (mixed domains: mathematics, science, code, tool-integrated reasoning with domain-specific judges), asynchronous Agentic RL (decoupled rollout engines via a Multi-Task Rollout Orchestrator, TITO gateway, direct double-sided importance sampling, DP-aware routing, >10K verifiable SWE/terminal/search environments), and General RL (foundational correctness, emotional intelligence, task-specific quality with hybrid rule-based/ORM/GRM rewards and human-authored anchors); on-policy cross-stage distillation as final stage[29]
Distillationon-policy cross-stage distillation: final checkpoints of preceding SFT/RL stages serve as teachers, with the advantage term replaced by the log-ratio between teacher and student policies (group size 1, batch 1024)[30]

Modality

Typetext-only
Attachmentn/a[31]

Inference efficiency

KV-cache designMLA low-rank latent KV cache (576-dim per token: KV LoRA 512 + 64); DSA lightning indexer retrieves top-k = 2048 KV entries per query so attention is computed sparsely over cached entries; multi-node DP-attention used in RL rollouts to avoid copying KV across ranks[32]
Quantisation shippedINT4 QAT from the SFT stage with bitwise-identical training/inference kernel (§2.4.3); on Ascend: W4A8 mixed-precision (MoE experts W4A8/INT4, attention and MLP blocks W8A8/INT8, QuaRot + Flex_AWQ_SSZ calibration) fitting the 750B model on one Atlas 800T A3 (§5); FP8 for RL rollout inference (§3.6.2)[33]
Speculative decodingMulti-Token Prediction with parameter sharing: 3 MTP layers share parameters during training (memory cost of a single MTP layer, matching DeepSeek-V3) while increasing acceptance rate; accept length 2.76 vs DeepSeek-V3.2's 2.55 at 4 speculative steps; MTP also accelerates RL rollouts[34]
Serving optimisationsslime RL infrastructure (mixed-precision rollouts, MTP, Prefill-Decode disaggregation, no-queue multi-node serving with DP-attention, heartbeat-driven fault tolerance); DP-aware routing with consistent hashing for KV-cache locality; on Ascend: fused Lightning Indexer / Sparse Flash Attention / MLAPO kernels, vLLM-Ascend and SGLang with asynchronous scheduling, RadixCache and prefix cache, FlashComm[35]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
HLE30.5Table 7
HLE (w/ Tools)50.4Table 7
AIME 2026 I92.7Table 7
HMMT Feb. 202597.9Table 7
IMO-AnswerBench82.5Table 7
GPQA-Diamond86.0Table 7
LongBench v264.5Table 7
SWE-bench Verified77.8Table 7
SWE-bench Multilingual73.3Table 7
Terminal-Bench 2.0 (Terminus-2)56.2 / 60.7Table 7
CyberGym43.2Table 7
BrowseComp (w/ Context Manage)75.9Table 7
BrowseComp-ZH72.7Table 7
τ2-Bench89.7Table 7
MCP-Atlas (Public Set)67.8Table 7
Tool-Decathlon39.2Table 7
Vending-Bench 2$4,432Table 7
GDPval-AA Elo1409Table 7
Artificial Analysis Intelligence Index v4.050§1
SWE-rebench (Resolved Rate)42.1Table 9
CC-Bench-V2 Backend Engineering (Pass@1)25.8Table 8
CC-Bench-V2 Long-horizon Chained Tasks (Pass@1)52.3Table 8

Provenance references

  1. stated · Title
  2. stated · Title page byline
  3. stated · pipeline
  4. stated · Abstract ('Code, models, and more information are available at https://github.com/zai-org/GLM-5'); §1 ('We release GLM-5 to the community'); §1 (first open weights model to score 50 on the AA Intelligence Index v4.0); §7 ('By open-sourcing the model')
  5. stated · §2.1 ('a 744B parameter model'), Table 10 (# Total Parameters 744B)
  6. stated · §2.1 ('40B active parameters'), Table 10 (# Activated Parameters 40B)
  7. stated · Table 10 (Hidden Dim 6144)
  8. stated · Table 10 (MoE Intermediate Dim 2048)
  9. stated · Table 10 (# Attention Heads 64)
  10. n/a · §2.1, Table 10 (# Key-Value Heads '-')
  11. stated · Table 10 (QK Head Dim 192)
  12. stated · Table 10 (Vocabulary Size 154880)
  13. stated · §2.1 (Multi-latent Attention), Table 10 (Q LoRA Dim 2048, KV LoRA Dim 512)
  14. stated · §2.1.1, §3.2 ('DSA is lossless by construction ... enabling application to all layers'), Table 10 (# Indexer Attn Heads 32, # Indexer Head Dim 128)
  15. stated · Table 10 (# Routed Experts 8)
  16. stated · Table 10 (# Shared Experts 1)
  17. stated · §2.3, §2.1.1
  18. stated · §2.3 (200K stage), §3.1 (SFT extends max context to 202,752 tokens)
  19. stated · §2.3 (mid-training extends the context window across 32K/128K/200K stages), §3.1 (SFT extends maximum context length to 202,752 tokens), §2.1.1 (DSA warm-up trains 202,752-token sequences)
  20. stated · Appendix B.2 (SWE-bench evaluated 'with a 200K context window'), §6.1.1 (max generation 202,752 tokens for HLE-with-tools)
  21. stated · §2 ('totaling 28.5 trillion tokens for the base model'), §1 ('we extend the training token budget to 28.5T tokens')
  22. stated · §2.2
  23. stated · §2, §2.3, §2.1.1
  24. stated · §2.1, §2.4.1, Appendix A ('we follow the setting of GLM-4.5, including the Muon optimizer')
  25. stated · Appendix A
  26. stated · §2.4.1, §2.4.2
  27. stated · §3.1 (multi-task SFT over General Chat, Reasoning, and Coding & Agent categories; INT4 QAT from the SFT stage; max context 202,752 tokens; interleaved/preserved/turn-level thinking modes)
  28. stated · §3.2 ('Our RL algorithm builds upon GRPO [40] and incorporates the IcePop technique', group-normalized advantages, PPO-style clipped importance ratios, β=2, ϵlow=0.2, ϵhigh=0.28); §4.1 (agentic RL uses a group-wise policy optimization objective over K sampled traces)
  29. stated · §3.2, §3.3, §3.4, §3.5, §4
  30. stated · §3.5
  31. n/a · §1, §6 (text-only model)
  32. stated · §2.1, §3.2, §3.6.2
  33. stated · §2.4.3, §5, §3.6.2
  34. stated · §2.1 (Table 2), §3.6.2
  35. stated · §3.6, §4.1.2, §5