The technical report for GLM-5, successor to GLM-4.5.
The stated goal
GLM-5 is "designed to transition the paradigm of vibe coding to agentic engineering" — a shift from generating code interactively toward executing sustained engineering work autonomously. It builds on the ARC (agentic, reasoning, coding) capabilities of its predecessor.
Technical contributions
DSA is adopted "to significantly reduce training and inference costs while maintaining long-context fidelity" — the cost-per-capability improvement that underwrites the open-weight economics discussed at Inference Economics and Token Pricing.
Asynchronous RL infrastructure. The report describes "a new asynchronous reinforcement learning infrastructure that drastically improves post-training efficiency by decoupling generation from training." In synchronous RL, the training loop waits on rollout generation; decoupling them removes that bottleneck, which matters disproportionately for long-horizon agentic tasks where individual rollouts are slow.
Asynchronous agent RL algorithms. Beyond the infrastructure, the report proposes "novel asynchronous agent RL algorithms that further improve RL quality, enabling the model to learn from complex, long-horizon interactions more" effectively — the algorithmic counterpart to the infrastructure change, aimed at the same long-horizon target.
Together these are the training-side answer to the capability gap that Mowshowitz and UK AISI both locate in long-horizon rather than narrow tasks.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| IMO-AnswerBench | GLM-5 | 82.5% | max generation length 131,072 tokens, temperature 1.0, top_p 0.95; developer-reported | 6.1 (Table 7); B.2 |
| IMO-AnswerBench | GLM-4.7 | 82.0% | max generation length 131,072 tokens, temperature 1.0, top_p 0.95; developer-reported | 6.1 (Table 7); B.2 |
| IMO-AnswerBench | DeepSeek-V3.2 | 78.3% | max generation length 131,072 tokens, temperature 1.0, top_p 0.95; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| IMO-AnswerBench | Kimi K2.5 | 81.8% | max generation length 131,072 tokens, temperature 1.0, top_p 0.95; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| IMO-AnswerBench | Claude Opus 4.5 | 78.5% | max generation length 131,072 tokens, temperature 1.0, top_p 0.95; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| IMO-AnswerBench | Gemini 3 Pro | 83.3% | max generation length 131,072 tokens, temperature 1.0, top_p 0.95; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| IMO-AnswerBench | GPT-5.2 (xhigh) | 86.3% | max generation length 131,072 tokens, temperature 1.0, top_p 0.95; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| LongBench v2 | GLM-5 | 64.5% | setting not stated beyond the general reasoning configuration; developer-reported | 6.1 (Table 7); B.2 |
| LongBench v2 | GLM-4.7 | 59.1% | setting not stated beyond the general reasoning configuration; developer-reported | 6.1 (Table 7); B.2 |
| LongBench v2 | DeepSeek-V3.2 | 59.8% | setting not stated beyond the general reasoning configuration; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| LongBench v2 | Kimi K2.5 | 61.0% | setting not stated beyond the general reasoning configuration; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| LongBench v2 | Claude Opus 4.5 | 64.4% | setting not stated beyond the general reasoning configuration; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| LongBench v2 | Gemini 3 Pro | 68.2% | setting not stated beyond the general reasoning configuration; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| LongBench v2 | GPT-5.2 (xhigh) | 59.8% | setting not stated beyond the general reasoning configuration; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| BrowseComp-ZH | GLM-5 | 72.7% | setting not stated; developer-reported | 6.1 (Table 7); B.2 |
| BrowseComp-ZH | GLM-4.7 | 66.6% | setting not stated; developer-reported | 6.1 (Table 7); B.2 |
| BrowseComp-ZH | DeepSeek-V3.2 | 65.0% | setting not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| BrowseComp-ZH | Kimi K2.5 | 62.3% | setting not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| BrowseComp-ZH | Claude Opus 4.5 | 62.4% | setting not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| BrowseComp-ZH | Gemini 3 Pro | 66.8% | setting not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| BrowseComp-ZH | GPT-5.2 (xhigh) | 76.1% | setting not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| Tool-Decathlon | GLM-5 | 39.2% | setting not stated; developer-reported | 6.1 (Table 7); B.2 |
| Tool-Decathlon | GLM-4.7 | 23.8% | setting not stated; developer-reported | 6.1 (Table 7); B.2 |
| Tool-Decathlon | DeepSeek-V3.2 | 35.2% | setting not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| Tool-Decathlon | Kimi K2.5 | 27.8% | setting not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| Tool-Decathlon | Claude Opus 4.5 | 43.5% | setting not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| Tool-Decathlon | Gemini 3 Pro | 36.4% | setting not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| Tool-Decathlon | GPT-5.2 (xhigh) | 46.3% | setting not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.1 (Table 7); B.2 |
| CC-Bench-V2 Frontend — HTML ISR | GLM-5 | 38.9% | Instance Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — HTML ISR | GLM-4.7 | 35.4% | Instance Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — HTML ISR | Claude Opus 4.5 | 52.2% | Instance Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — HTML CSR | GLM-5 | 76.3% | Check-item Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — HTML CSR | GLM-4.7 | 64.9% | Check-item Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — HTML CSR | Claude Opus 4.5 | 82.2% | Check-item Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — React ISR | GLM-5 | 34.6% | Instance Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — React ISR | GLM-4.7 | 17.2% | Instance Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — React ISR | Claude Opus 4.5 | 39.7% | Instance Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — React CSR | GLM-5 | 71.0% | Check-item Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — React CSR | GLM-4.7 | 49.4% | Check-item Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — React CSR | Claude Opus 4.5 | 70.7% | Check-item Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — Vue ISR | GLM-5 | 32.7% | Instance Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — Vue ISR | GLM-4.7 | 24.5% | Instance Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — Vue ISR | Claude Opus 4.5 | 46.9% | Instance Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — Vue CSR | GLM-5 | 77.1% | Check-item Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — Vue CSR | GLM-4.7 | 53.8% | Check-item Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend — Vue CSR | Claude Opus 4.5 | 74.3% | Check-item Success Rate; internal CC-Bench-V2 frontend; Agent-as-a-Judge (Claude Code with Claude Sonnet 4.5 and Playwright MCP) over built projects; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — React BSR | GLM-5 | 100% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — React BSR | GLM-4.7 | 65.0% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — React BSR | Claude Opus 4.5 | 95.0% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — Vue BSR | GLM-5 | 100% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — Vue BSR | GLM-4.7 | 70.0% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — Vue BSR | Claude Opus 4.5 | 100% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — Svelte BSR | GLM-5 | 100% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — Svelte BSR | GLM-4.7 | 60.0% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — Svelte BSR | Claude Opus 4.5 | 90.0% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — Next.js BSR | GLM-5 | 95.0% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — Next.js BSR | GLM-4.7 | 70.0% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; developer-reported | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — Next.js BSR | Claude Opus 4.5 | 80.0% | Build Success Rate (share of generated projects that initialize and run); internal CC-Bench-V2; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8) |
| CC-Bench-V2 Frontend Build — Overall BSR | GLM-5 | 98.0% | overall Build Success Rate stated in text; developer-reported | 6.2.1 |
| CC-Bench-V2 Backend Engineering — Pass@1 | GLM-5 | 25.8% | 85 tasks in six languages; task solved only if all human-written unit tests pass; internal CC-Bench-V2; developer-reported | 6.2 (Table 8); 6.2.2 |
| CC-Bench-V2 Backend Engineering — Pass@1 | GLM-4.7 | 19.6% | 85 tasks in six languages; task solved only if all human-written unit tests pass; internal CC-Bench-V2; developer-reported | 6.2 (Table 8); 6.2.2 |
| CC-Bench-V2 Backend Engineering — Pass@1 | Claude Opus 4.5 | 26.9% | 85 tasks in six languages; task solved only if all human-written unit tests pass; internal CC-Bench-V2; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8); 6.2.2 |
| CC-Bench-V2 Long-horizon: Repo Exploration — Pass@1 | GLM-5 | 65.6% | pass@1 averaged over three runs; solved if the agent reads the target file; internal CC-Bench-V2; developer-reported | 6.2 (Table 8); 6.2.3 |
| CC-Bench-V2 Long-horizon: Repo Exploration — Pass@1 | GLM-4.7 | 47.8% | pass@1 averaged over three runs; solved if the agent reads the target file; internal CC-Bench-V2; developer-reported | 6.2 (Table 8); 6.2.3 |
| CC-Bench-V2 Long-horizon: Repo Exploration — Pass@1 | Claude Opus 4.5 | 64.5% | pass@1 averaged over three runs; solved if the agent reads the target file; internal CC-Bench-V2; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8); 6.2.3 |
| CC-Bench-V2 Long-horizon: Chained Tasks — Pass@1 | GLM-5 | 52.3% | pass@1 on individual tasks within multi-step task chains mined from merged pull requests; internal CC-Bench-V2; developer-reported | 6.2 (Table 8); 6.2.3 |
| CC-Bench-V2 Long-horizon: Chained Tasks — Pass@1 | GLM-4.7 | 43.0% | pass@1 on individual tasks within multi-step task chains mined from merged pull requests; internal CC-Bench-V2; developer-reported | 6.2 (Table 8); 6.2.3 |
| CC-Bench-V2 Long-horizon: Chained Tasks — Pass@1 | Claude Opus 4.5 | 61.6% | pass@1 on individual tasks within multi-step task chains mined from merged pull requests; internal CC-Bench-V2; reported by Zhipu AI / Z.ai in the GLM-5 technical report | 6.2 (Table 8); 6.2.3 |
| SWE-rebench — January 2026, resolved rate | Claude Opus 4.6 | 52.9% | resolved rate, SEM ±1.06%; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, pass@5 | Claude Opus 4.6 | 70.8% | pass@5; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, resolved rate | GPT-5.2 (xhigh) | 51.7% | resolved rate, SEM ±1.21%; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, pass@5 | GPT-5.2 (xhigh) | 58.3% | pass@5; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, resolved rate | Claude Sonnet 4.5 | 47.1% | resolved rate, SEM ±1.69%; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, pass@5 | Claude Sonnet 4.5 | 60.4% | pass@5; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, resolved rate | Gemini 3 Pro | 46.7% | resolved rate, SEM ±2.04%; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, pass@5 | Gemini 3 Pro | 58.3% | pass@5; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, resolved rate | Claude Opus 4.5 | 43.8% | resolved rate, SEM ±0.93%; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, pass@5 | Claude Opus 4.5 | 58.3% | pass@5; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, resolved rate | GLM-5 | 42.1% | resolved rate, SEM ±1.21%; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, pass@5 | GLM-5 | 50.0% | pass@5; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, resolved rate | GLM-4.7 | 41.3% | resolved rate, SEM ±2.12%; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, pass@5 | GLM-4.7 | 56.3% | pass@5; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, resolved rate | Kimi K2.5 | 37.9% | resolved rate, SEM ±1.21%; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| SWE-rebench — January 2026, pass@5 | Kimi K2.5 | 50.0% | pass@5; SWE-rebench January 2026 window; official SWE-rebench leaderboard figures reproduced by Z.ai in the GLM-5 technical report | 6.2.4 (Table 9) |
| BBH | DeepSeek-V3 (base) | 88.4% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| BBH | Kimi K2 (base) | 88.7% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| BBH | GLM-4.5 (base) | 86.2% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| BBH | GLM-5 (base) | 87.4% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| HellaSwag | DeepSeek-V3 (base) | 88.9% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| HellaSwag | Kimi K2 (base) | 94.6% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| HellaSwag | GLM-4.5 (base) | 87.1% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| HellaSwag | GLM-5 (base) | 88.1% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| PIQA | DeepSeek-V3 (base) | 84.7% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| PIQA | GLM-4.5 (base) | 85.3% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| PIQA | GLM-5 (base) | 84.6% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| TriviaQA | DeepSeek-V3 (base) | 82.9% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| TriviaQA | Kimi K2 (base) | 85.1% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| TriviaQA | GLM-4.5 (base) | 80.0% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| TriviaQA | GLM-5 (base) | 80.9% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| EvalPlus | DeepSeek-V3 (base) | 65.6% | base model; Pass@1; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| EvalPlus | Kimi K2 (base) | 80.3% | base model; Pass@1; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| EvalPlus | GLM-4.5 (base) | 78.1% | base model; Pass@1; number of shots not stated; developer-reported | B.1 (Table 11) |
| EvalPlus | GLM-5 (base) | 87.0% | base model; Pass@1; number of shots not stated; developer-reported | B.1 (Table 11) |
| LiveCodeBench-Base | DeepSeek-V3 (base) | 24.6% | base model; Pass@1; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| LiveCodeBench-Base | Kimi K2 (base) | 26.3% | base model; Pass@1; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| LiveCodeBench-Base | GLM-4.5 (base) | 28.1% | base model; Pass@1; number of shots not stated; developer-reported | B.1 (Table 11) |
| LiveCodeBench-Base | GLM-5 (base) | 34.4% | base model; Pass@1; number of shots not stated; developer-reported | B.1 (Table 11) |
| CLUEWSC | DeepSeek-V3 (base) | 82.7% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| CLUEWSC | GLM-4.5 (base) | 83.5% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| CLUEWSC | GLM-5 (base) | 84.2% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| C-Eval | DeepSeek-V3 (base) | 90.1% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| C-Eval | Kimi K2 (base) | 92.5% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| C-Eval | GLM-4.5 (base) | 86.9% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| C-Eval | GLM-5 (base) | 88.8% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| C3 | DeepSeek-V3 (base) | 78.6% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| C3 | GLM-4.5 (base) | 83.1% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| C3 | GLM-5 (base) | 80.3% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| Chinese-SimpleQA | DeepSeek-V3 (base) | 72.1% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| Chinese-SimpleQA | Kimi K2 (base) | 77.6% | base model; EM; number of shots not stated; reported by Zhipu AI / Z.ai in the GLM-5 technical report | B.1 (Table 11) |
| Chinese-SimpleQA | GLM-4.5 (base) | 70.1% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
| Chinese-SimpleQA | GLM-5 (base) | 74.6% | base model; EM; number of shots not stated; developer-reported | B.1 (Table 11) |
Relationships
- supports: GLM-5 and GLM-5.1 — the technical documentation for the release
- related: GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models (Zhipu AI, 2025) — the predecessor report
- related: GLM-5.2, Open-Weight Frontier Models, Inference Economics and Token Pricing, Zhipu AI