RECENTLY UPDATED / (SHANGHAI TIME)
LIN YUEJI / ARCHIVE
Explore GLM-5.2 evaluations, coding agents, long-context workflows, integrations, and deployment lessons.
USE CASE MAP / 01
This collection preserves the key result, evidence type, creator, and original source for 258 public cases so you can quickly spot repeatable methods.
HOW TO USE / 02
Narrow by task, verify context at the original source, then reproduce the most useful path through EvoLink.
Start with coding, agents, creative work, evaluation, or another task family.
Use titles, methods, or creators to find the closest case to your current goal.
Select the creator to review the full demo, limits, context, and original explanation.
Choose a method worth testing, connect to the model, and turn it into your workflow.
REAL CREATOR CASES / 03
Up to four cases per row. Select a creator to open the original case; images and videos load only when needed.
258 cases
Use this case to benchmark fine-tuned local GLM-5.2 tool use rather than raw API baselines, because volatilemarkts says a 753GB FP8 fine-tune plus a custom FP16 indexer raised SeraphimSerapis/tool-eval-bench from 83 percent on the standard GLM 5.2 API to 94 percent.
Use this case to benchmark GLM-5.2 on real code-audit harnesses instead of chat demos, because Aikido says its AI Code Analysis benchmark on 26 known CVEs found GLM-5.2 rediscovering 16 at pass@3 and gaining three more findings at max reasoning for only about 1.3x the cost.
Use this case to evaluate GLM-5.2 on public-equity research agents, because karinanguyen says DiligenceBench placed GLM 5.2 near the top and showed that the finance harness can make strong models both better and cheaper.
Use this case to benchmark GLM-5.2 on physics-heavy single-file browser builds, because AlicanKiraz0 says GLM 5.2 Max won a Gargantua geodesic-raytracer task by balancing numerical correctness and real-time rendering discipline better than the peer models tested.
Use this case to budget GLM-5.2 for long-horizon benchmark workloads, because Artificial Analysis says GLM-5.2 Max averaged about 43K output tokens per Intelligence Index task versus 25K for Inkling and lower totals for Kimi K2.6 and DeepSeek v4 Pro Max.
Use this case to test a verifier-gated two-model coding route, because gmicloud says Opus 4.8 first plus GLM 5.2 FP8 as rescue solved 94 of 100 frozen EvalPlus tasks, five more than Fable 5, at about 47 percent lower cost.
Use this case to compare GLM-5.2 on algorithm-heavy browser physics builds, because AlicanKiraz0 ran a Stable Fluids HTML benchmark and scored GLM 5.2 Max at 88 out of 100 while costing about $1.17, ahead of Opus 4.8 and Fable 5 but behind GPT 5.6 Sol.
Use this case to place GLM-5.2 on a long-horizon capability curve, because Epoch AI estimates a score of 152 on its Capabilities Index and calls it the highest open-weight model in its evaluation set.
Use this case to benchmark GLM-5.2 on a large private engineering codebase, because Databricks says its internal eval over work from 3,000-plus engineers found GLM 5.2 performed extremely well and that harness choice alone can cut cost by about 2x.
Use this case to benchmark GLM-5.2 on scientific-agent work, because NatureBench says GLM-5.2 debuted at number two overall and took the open-weight lead across 90 tasks in six scientific domains.
Use this case to compare GLM-5.2 against GPT-5.5 on the same agent harness, because a 45-task Terminal-Bench run put GLM-5.2 at 25 wins versus GPT-5.5 at 29 while costing about 40% less with prompt caching.
Use this case to benchmark GLM-5.2 on real legal-agent work, because Harvey LAB-AA puts GLM-5.2 Max at a 7.5% all-pass rate, tied with Claude Opus 4.8 on 120 private tasks across 24 practice areas.
Use this case to compare GLM-5.2 on business-rule SaaS automation instead of coding-only benchmarks, because Artificial Analysis reports GLM-5.2 Max at 27.8% and calls it the leading open-weights model on AutomationBench-AA.
Use this case to compare GLM-5.2 on numerical-physics coding benchmarks, because AlicanKiraz0 ran a chaotic three-body simulator task and gave GLM 5.2 Max the top score at 91 out of 100.
Use this case to track GLM-5.2 on agentic game-development benchmarks, because GameDevBench expanded to 333 tasks and says GLM-5.2 is now the strongest open-source model on its leaderboard despite lacking vision.
Use this case to compare GLM-5.2 on a constrained Cursor coding benchmark, because AlicanKiraz0 ran six models on an HTML double-pendulum simulator and scored GLM 5.2 Max at 88 out of 100, behind Fable and Sonnet but ahead of GPT-5.5, Kimi K2.7 Code, and a failed Composer run.
Use this case to compare GLM-5.2 on real post-cutoff engineering tasks where cost matters as much as score, because Morgan Linton says VulcanBench gave GLM 5.2 High, Fable 5 Low, and Sonnet 5 High the same 80 percent score across 10 repos while GLM landed in the middle on cost.
Use this case to track GLM-5.2 on a continuously updated SWE agent leaderboard, because the latest SWE rebench post reports 51.1 percent with 2.62 million tokens, clearly ahead of the newly added DeepSeek, MiMo, Qwen, and Gemma runs.
Use this case to test GLM-5.2 on business-tool agent work instead of chat-only evals, because Composio reports 40 out of 41 on GitHub, Jira, and LaunchDarkly tasks and says GLM was the only model to catch a pending-approval edge case.
Use this case to measure GLM-5.2 on offensive-security-style bug finding and patching, because CyberBench puts it second overall on 60 real OSS-Fuzz vulnerabilities.
Use the Artificial Analysis post to compare GLM-5.2 against other open-weight and proprietary frontier models on intelligence and cost per task.
Use this case to evaluate GLM-5.2 on real front-end coding tasks judged by arena-style comparisons.
Use this case to judge whether GLM-5.2 can handle design-plus-code tasks rather than only text-heavy coding benchmarks.
Use the FrontierSWE post to compare GLM-5.2 against GPT-5.5, Opus, and Fable-style models on software-engineering tasks.
Use the DeepSWE case to understand GLM-5.2 as a strong open model for difficult software-engineering evaluation tasks.
Use this case when evaluating GLM-5.2 for terminal-oriented coding and agent workflows.
Use this SWELancer case as a concrete multi-metric comparison between GLM-5.2 and GPT-5.5 on task success, reward, and completion time.
Use this case to inspect GLM-5.2 on grounded multi-step reasoning rather than only coding leaderboards.
Use this case to compare GLM-5.2 with closed frontier models on grounded reasoning tasks.
Use this case when checking whether benchmark gains come from valid implementation behavior instead of shortcutting.
Use this case as a fast signal for open-weight model progress on game-like benchmark tasks.
Use this case to evaluate latency-sensitive workflows where speed matters alongside intelligence.
Use this case to evaluate GLM-5.2 on GPU-kernel coding across KernelBench-Hard and KernelBench-Mega, where open agent traces make the result inspectable.
Use this case to track GLM-5.2 on DeepSWE at max effort, where the posted leaderboard puts it first among open models with a 44% pass@1 score.
Use this case to evaluate GLM-5.2 beyond coding tasks on adversarial multi-turn debate, where the max-reasoning variant placed second behind Claude models.
Use this case to compare GLM-5.2 on uncertainty handling, where the posted AA-Omniscience result shows a lower hallucination rate than several other frontier models.
Use this case to compare GLM-5.2 on long-horizon knowledge work rather than coding-only leaderboards.
Use this case to judge GLM-5.2 on game-building quality, where the model reached second place on Game Dev Arena and became the top open-weight lab in that ranking.
Use this case to compare GLM-5.2 Max on post-training agent reliability, not just headline score, because the leaderboard also reports zero failed runs across 84 tasks.
Use this case to judge GLM-5.2 on real private-repo engineering tasks instead of only public benchmarks, because the reported win includes score, speed, and cost per task.
Use this case to compare GLM-5.2 on long-horizon knowledge-work tasks where time per task matters alongside benchmark score.
Use this case to inspect GLM-5.2's frontend edge through pairwise head-to-head results instead of relying on a single rank screenshot.
Use this case to track GLM-5.2 across codebase QnA, test writing, and refactoring rather than only single-task SWE leaderboards.
Use this case to route GLM-5.2 inside a Codex-centered coding loop instead of staying locked to one model, because vista8 says OpenCodex lets the same environment switch between GLM 5.2, Kimi K3, GPT-5.6 Sol, and Grok 4.5 for frontend design, backend work, and live X search.
Use this case to structure a role-based multi-agent lab around GLM-5.2 instead of one monolithic assistant, because MichaelGannotti says an 11-agent Hermes setup routes tasks dynamically across DGX Spark, Ryzen workstations, and cloud models including GLM 5.2 for software, research, marketing, and coordination work.
Use this case to validate a self-hosted GLM-5.2 coding agent against the official route, because dangerm00se says a Hermes plus GLM-5.2 hybrid on 4x RTX 6000 PCIe matched 59 of 60 tasks from the official API while delivering 3,149 tok/s prefill, 0.37s warm TTFT, and 35.9 tok/s decode.
Use this case to evaluate a local-first GLM-5.2 coding agent, because chenzeling4 says LM Studio Bionic pairs GLM 5.2 with local document sandboxes, inline code diffs, rollback checkpoints, and on-device voice transcription.
Use this case to compare first-pass web-dev quality instead of raw completion speed, because Lumenix0 says GLM 5.2 in Claude Code beat GPT 5.5 in Codex on design quality and functional completeness across three real tasks.
Showing 48 / 258
BUILD WITH EVOLINK / 04
Connect to the model through EvoLink and build from directions already demonstrated in public.
