We present GLM-5, a next-generation foundation model designed to transition the paradigm of vibe coding to agentic engineering. Building upon the agentic, reasoning, and coding (ARC) capabilities of its predecessor, GLM-5 adopts DSA to significantly reduce training and inference costs while maintaining long-context fidelity. To advance model alignment and autonomy, we implement a new asynchronous reinforcement learning infrastructure that drastically improves post-training efficiency by decoupling generation from training. Furthermore, we propose novel asynchronous agent RL algorithms that further improve RL quality, enabling the model to learn from complex, long-horizon interactions more effectively. Through these innovations, GLM-5 achieves state-of-the-art performance on major open benchmarks. Most critically, GLM-5 demonstrates unprecedented capability in real-world coding tasks, surpassing previous baselines in handling end-to-end software engineering challenges. Code, models, and more information are available at https://github.com/zai-org/GLM-5.
Figures & tables
Figure 1: Results of GLM-5, DeepSeek-V3.2, Claude Opus 4.5, Gemini 3 Pro, and GPT-5.2 (xhigh) on 8 agentic, reasoning, and coding benchmarks: Humanity’s Last Exam, SWE-bench Verified, SWE-bench Multilingual, Terminal-Bench 2.0, BrowseComp, MCP-Atlas, τ2 -Bench, Vending Bench 2.
Figure 3: On LMArena, GLM-5 is the #1 open model in both Text Arena and Code Arena.
Figure 4: Results on several long-horizon tasks. Left: Vending-Bench 2; Right: CC-Bench-V2.
Figure 5: Overall training pipeline of GLM-5.
Dataset
Hellaswag
MMLU
C-Eval
RACE
BBH
GSM8K
HumanEval
GQA-8
77.3
61.2
60.0
79.6
53.3
47.6
38.5
MLA
77.3
61.5
59.7
77.8
48.9
46.2
33.5
MLA + Muon Split
77.8
62.5
62.1
79.9
51.8
45.0
36.7
MLA-256 + Muon Split
77.4
62.0
59.9
79.6
51.3
47.5
36.6
Table 1: Evaluation results for GQA-8 and variants of MLA.
Model
Accept Length
DeepSeek-V3.2
2.55
GLM-5
2.76
Table 2: Comparison of accept lengths of DeepSeek-V3.2 and GLM-5.
MQ-NIAH-128k
MV-NIAH-128k
SQuAD-128k
HotpotQA-128k
MLA
100.0
95.5
79.7
66.3
DSA
100.0
97.0
86.0
63.0
Table 3: Comparison of long-context benchmarks between MLA and DSA base models.
Figure 6: SFT loss curves comparison between MLA and DSA training. Results are smoothed by Running Average with a window size of 50.
4K
8K
16K
32K
64K
128K
GLM-9B (Full Attn)
95.19
93.67
92.01
91.09
85.35
75.28
SWA Interleave
94.87
54.02
25.89
12.61
8.32
6.51
SWA Pattern
95.78
92.54
88.92
82.52
70.23
53.95
Table 4: RULER benchmark results for the GLM-9B baseline and two SWA variants without any additional training . Both SWA methods use a 1:1 ratio of full-attention to SWA layers with a 4096-token window size. The search-based SWA pattern is discovered once at 16k context length and applied uniformly across all input lengths.
RULER (64K / 128K)
MRCR (64K / 128K)
HELMET-ICL (64K / 128K)
RepoQA (64K / 128K)
GLM-9B
85.35 / 75.28
36.53 / 35.39
77.68 / 77.36
69.00 / 65.83
SWA Interleave
65.94 / 44.93 ( ↓ 19.41 / ↓ 30.35)
30.03 / 28.83 ( ↓ 6.50 / ↓ 6.56)
75.96 / 63.52 ( ↓ 1.72 / ↓ 13.84)
50.33 / 39.33 ( ↓ 18.67 / ↓ 26.50)
SWA Pattern
83.72 / 69.59 ( ↓ 1.63 / ↓ 5.69)
35.02 / 33.58 ( ↓ 1.51 / ↓ 1.81)
76.48 / 74.60 ( ↓ 1.20 / ↓ 2.76)
62.33 / 51.17 ( ↓ 6.67 / ↓ 14.66)
GDN
76.76 / 64.00 ( ↓ 8.59 / ↓ 11.28)
31.72 / 30.22 ( ↓ 4.81 / ↓ 5.17)
76.88 / 74.84 ( ↓ 0.80 / ↓ 2.52)
65.50 / 56.17 ( ↓ 3.50 / ↓ 9.66)
SimpleGDN
81.76 / 67.03 ( ↓ 3.59 / ↓ 8.25)
33.03 / 31.27 ( ↓ 3.50 / ↓ 4.12)
79.80 / 81.84 ( ↑ 2.12 / ↑ 4.48)
65.50 / 58.50 ( ↓ 3.50 / ↓ 7.33)
Table 5: Long-context benchmark results. All efficient attention variants are continual-trained from the GLM-9B full-attention baseline. SWA pattern denotes search-based layer selection; SWA interleave denotes the fixed alternating pattern. Δ @64K and Δ @128K show the difference relative to the full-attention baseline at 64K and 128K context lengths, respectively.
4K
8K
16K
32K
64K
128K
GLM-4.7-Flash
97.44
96.72
95.83
92.96
85.34
79.21
GLM-4.7-Flash + DSA warmup
97.51
96.54
95.40
90.09
84.05
71.35
GLM-4.7-Flash + DSA
96.77
96.25
96.69
93.45
87.06
78.86
Table 6: RULER benchmark results for the GLM-4.7-Flash with DSA. The warmup-only variant trains only the indexer while keeping the base model frozen, the full DSA variant jointly trains both for 150B tokens.
Figure 7: Illustration of Interleaved Thinking and Preserved Thinking.
Figure 8: Accuracy of BrowseComp with different context management strategies from GLM-4.7 (gray baselines) to GLM-5 (colored strategies).
Figure 9: Examples of reward hacking in the slides RL training. Our runtime rendering obtains grounded attribute values, making the evaluation robust to such hacking behaviors.
GLM-5
GLM-4.7
DeepSeek -V3.2
Kimi K2.5
Claude Opus 4.5
Gemini 3 Pro
GPT-5.2 (xhigh)
Reasoning & General
HLE
30.5
24.8
25.1
31.5
28.4
37.2
35.4
HLE (w/ Tools)
50.4
42.8
40.8
51.8
43.4*
45.8*
45.5*
AIME 2026 I
92.7
92.9
92.7
92.5
93.3
90.6
-
HMMT Feb. 2025
97.9
97.1
92.5
95.4
92.9
97.3
99.4
HMMT Nov. 2025
96.9
93.5
90.2
91.1
91.7
93.0
97.1
Table 7: Comparison between GLM-5 and open-source/proprietary models. Results marked with * are from the full set of HLE. Results marked with † are evaluated on a verified version of Terminal-Bench 2.0, fixing some ambiguous instructions. The GDPval-AA Elo scores are recorded on 15th Feb., 2026. The highest score for each benchmark is bolded , and the second highest is underlined .
Figure 10: Agent-as-a-Judge evaluation pipeline. Each generated frontend project is first built to verify static correctness. Successfully built instances are then interactively tested by an autonomous Judge Agent, which determines the functional correctness of each check item.
Model
Resolved Rate (%)
Resolved Rate SEM ( ± , %)
Pass@5 (%)
Claude Opus 4.6
52.9%
1.06%
70.8%
GPT-5.2 (xhigh)
51.7%
1.21%
58.3%
Claude Sonnet 4.5
47.1%
1.69%
60.4%
Gemini 3 Pro
46.7%
2.04%
58.3%
Claude Opus 4.5
43.8%
0.93%
58.3%
GLM-5
42.1%
1.21%
50.0%
Table 9: Performance on SWE-rebench, January 2026.
Figure 11: Performance comparison between GLM-4.7 and GLM-5 across five real-world general ability domains.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Model
GLM-4.5
GLM-5
# Total Parameters
355B
744B
# Activated Parameters
32B
40B
# Dense Layers
3
3
# MoE Layers
89
75
# MTP Layers
1
1
Hidden Dim
5120
6144
Appendix
Table 10: Model architecture of GLM-4.5 and GLM-5. When counting parameters, for all models we include the parameters of MTP layers but not word embeddings and the output layer.
Benchmark (Metric)
DeepSeek-V3
Kimi-K2
GLM-4.5
GLM-5
Base
Base
Base
Base
Architecture
MoE
MoE
MoE
MoE
# Activated Params
37B
32B
32B
40B
# Total Params
671B
1043B
355B
744B
English
SimpleQA (EM)
26.6
35.3
30.0
36.0
BBH (EM)
88.4
88.7
86.2
87.4
Appendix
Table 11: Comparison among GLM-5-Base, GLM-4.5-Base and other representative open-source base models.
Category
Description
# Tasks
# Checkitems
Business Systems
Enterprise/Personal data and process management.
42
167
Web Games
Interaction and entertainment-focused games.
40
163
SVG/Canvas
Graphics rendering and interactive visualizations.
32
166
Creative Tools
Content creation and online editing tools.
28
160
Showcase Pages
Visual expression and information presentation.
27
115
Forms & Tables
Structured data entry and processing.
26
93
Appendix
Table 12: Distribution of frontend application scenarios.
Category
Description
# Tasks
# Checkitems
HTML
Vanilla HTML/CSS/JS development.
113
490
React
Component-based framework development.
58
249
Vue
Vue 3 + Vite progressive solution.
49
210
Appendix
Table 13: Statistics of technology stacks and evaluation units.
We present KAT-Coder-V2, an agentic coding model developed by the KwaiKAT team at Kuaishou. KAT-Coder-V2 adopts a "Specialize-then-Unify" paradigm that decomposes agentic coding into five expert domains - SWE, WebCoding, Terminal, WebSearch, and General - each undergoing independent supervised fine-tuning and reinforcement learning, before being consolidated into a single model via on-policy distillation. We develop KwaiEnv, a modular infrastructure sustaining tens of thousands of concurrent sandbox instances, and scale RL training along task complexity, intent alignment, and scaffold generalization. We further propose MCLA for stabilizing MoE RL training and Tree Training for eliminating redundant computation over tree-structured trajectories with up to 6.2x speedup. KAT-Coder-V2 achieves 79.6% on SWE-bench Verified (vs. Claude Opus 4.6 at 80.8%), 88.7 on PinchBench (surpassing GLM-5 and MiniMax M2.7), ranks first across all three frontend aesthetics scenarios, and maintains strong generalist scores on Terminal-Bench Hard (46.8) and tau^2-Bench (93.9). Our model is publicly available at https://streamlake.com/product/kat-coder.
Reinforcement learning (RL) is becoming increasingly important for post-training large language models (LLMs). Previous RL pipelines for LLMs were mostly synchronous and batch-interleaved, which is inefficient for long-horizon agentic tasks. Recently, asynchronous RL has emerged as a more efficient alternative by updating the model as rollouts arrive. However, existing asynchronous RL systems often emphasize throughput, while leaving training stability and task effectiveness largely underexplored. For example, a key challenge is that group-wise sampling in the widely-used GRPO framework does not naturally fit asynchronous agentic training. In this paper, we present Single-rollout Asynchronous Optimization (SAO) to address the stability and off-policy challenges in asynchronous RL. To reduce off-policy effects and improve generalization, we replace group-wise sampling with single-rollout sampling, that is, using one rollout per prompt. We further improve this single-rollout strategy with practical value-model training designs. To improve optimization stability, we introduce a strict double-side token-level clipping strategy. SAO is able to train stably for one thousand steps and consistently outperform GRPO and its variants on agentic coding and reasoning benchmarks, such as SWE-Bench Verified, BeyondAIME, and IMOAnswerBench. We also demonstrate that single-rollout RL is particularly effective in a simulated online learning setting, where the model must adapt to changing evolving environments. To this end, SAO is successfully deployed in the agentic RL pipeline for training the open GLM-5.2 model (750B-A40B).
Zhenyu Hou, Yujiang Li, Jie Tang +1
Tsinghua University · Work done while ZH and YL interned at Z.AI.
We present IQuest-Q1, a sparse Mixture-of-Experts (MoE) agentic foundation model with 320B total parameters (15B activated). IQuest-Q1 is developed through a three-stage process: (1) stable and efficient pre-training, combining Stable SplitNorm, Muon, and muP with a two-stage curriculum whose second stage increases the coding-data mixture, preserving a wide stability margin between safe and optimal learning rates; (2) fine-grained mid-training, expanding the data mixture across domains and formats and progressively extending the context; and (3) multi-scaffold with multi-version post-training, combining SFT, domain-specialized RL, and MOPD across various scaffolds to integrate complementary training signals and strengthen agentic coding capabilities. Furthermore, model merging combines different experts and checkpoints from different stages while mitigating conflicts among their parameter updates. The resulting IQuest-Q1 demonstrates capabilities in both coding and general tasks. We release IQuest-Q1 as an open-source model to support community research and development.