GLM-5：Agentic Engineering

Overview

Who it's for

Developers and researchers who want an open-source model for agentic and complex systems engineering, or who want to compare open models against frontier closed models on long-horizon tasks.

What you'll learn

1. 01What GLM-5 is: a model family built for 'agentic engineering'

   GLM-5 is a model + deployment guide rather than a tutorial — the technical report is titled 'From Vibe Coding to Agentic Engineering'. Scaling from 355B (32B active) to 744B parameters (40B active) with pretraining data growing from 23T to 28.5T tokens, GLM-5 integrates DeepSeek Sparse Attention to cut deployment cost while preserving long-context capacity. Purpose-built for complex systems engineering and long-horizon agentic tasks, it significantly outperforms GLM-4.7 on CC-Bench-V2 and narrows the gap to Claude Opus 4.5.

2. 02How 'from Vibe Coding to Agentic Engineering' lands

   GLM-5.1's core selling point is not first-pass quality but sustained effectiveness on long-horizon tasks: it decomposes complex problems, runs experiments, reads results and pinpoints blockers, then revises its strategy through repeated iteration — sustaining optimization over hundreds of rounds and thousands of tool calls. The longer it runs, the better the result. GLM-5 ranks #1 among open-source models on Vending Bench 2 (running a simulated vending-machine business over a one-year horizon) with a final account balance of $4,432, approaching Claude Opus 4.5 and showing long-term planning and resource management.

3. 03Three generations, three jobs: GLM-5 / 5.1 / 5.2

   GLM-5 targets complex systems engineering and long-horizon agentic tasks; GLM-5.1 is the agentic-engineering flagship with SOTA on SWE-Bench Pro and a wide margin over GLM-5 on NL2Repo (repo generation) and Terminal-Bench 2.0; GLM-5.2 is the long-horizon flagship, delivering a solid 1M-token context for the first time, coding with multiple thinking-effort levels, and an IndexShare architecture that reuses one indexer across every four sparse-attention layers (2.9× fewer FLOPs per token at 1M context). On Terminal-Bench 2.1, 5.2 scores 81.0 (vs. 5.1's 62.0) — within a few points of closed-source frontier Claude Opus 4.8's 85.0.

4. 04Local deployment: multi-framework support + thinking-budget control

   Weights ship in BF16 and FP8 (744B-A40B) and run on SGLang, vLLM, Transformers, KTransformers, Unsloth and more, with Ascend NPU supported via vLLM-Ascend / xLLM / SGLang. The reasoning_effort parameter controls the thinking budget — max is the default, high must be passed explicitly, and enable_thinking=false turns thinking off entirely. Keep the default max to reproduce benchmark/leaderboard results.

5. 05Fine-tuning: asynchronous RL infra Slime + ms-swift

   The repo documents two fine-tuning paths: Slime is the GLM team's own asynchronous RL infrastructure, substantially improving training throughput and enabling finer-grained post-training iterations; ms-swift supports SFT, PPO and GRPO. For reproducing paper experiments or domain adaptation, official weights are on both Hugging Face and ModelScope.

Practice tips

• Download the weights (BF16 / FP8) from Hugging Face or ModelScope, serve locally with SGLang or vLLM, and test sustained output on one long-horizon agentic task.
• Compare GLM-5 / 5.1 / 5.2 side by side to see that first-pass accuracy and sustained long-horizon productivity are two different things.
• Tune reasoning_effort (max / high) and enable_thinking=false to measure the latency/quality trade-off at different thinking budgets; keep default max when reproducing benchmarks.
• Read the Vending Bench 2 and Terminal-Bench 2.1 results in the technical report to see how 'agentic engineering' differs from 'vibe coding' in evaluation design.

Companion workflows

Related resources