GLM-5: from Vibe Coding to Agentic Engineering

GLM-5 is Zhipu AI's open-weight flagship: a mixture-of-experts model with 744 billion parameters (40 billion active per token), trained on about 28.5 trillion tokens with context extended to 200K, using DeepSeek Sparse Attention to cut long-context cost. Benchmark tables put it near Claude Opus 4.5 and GPT-5.2 on coding and agent tasks.

The report's practical value is in how it trains agents. Agentic RL is run fully asynchronously: inference and training engines sit on different GPUs; the inference side keeps generating trajectories, and the trainer updates whenever enough have arrived, pushing new weights back periodically. That removes the idle time caused by a few very long rollouts, but it creates off-policy data, which the paper handles with a set of concrete mechanisms.

A token-in-token-out (TITO) gateway records the exact token IDs the inference engine produced, so the trainer never re-tokenizes text and misaligns actions and rewards. Direct double-sided importance sampling reuses the rollout's log-probabilities as the behavior policy and masks tokens whose ratio leaves a band. Samples are dropped if their oldest weight version is too stale, or if the sandbox crashed. A DP-aware router sends every turn of one rollout to the same data-parallel rank, so its prefix stays cached.

Environments are built at scale: over 10,000 verifiable software tasks from real issues in nine languages, thousands of synthesized terminal tasks, and multi-hop search questions. For long searches, a simple keep-recent-k rule replaces old tool results with a stub, and a hierarchical rule adds a full reset past 32K tokens.

Most teams will never train a 744B model. Several of GLM-5's mechanisms still apply to anyone running agents in production.

Pin a conversation to one cache. Multi-turn agents resend a growing prefix every turn. Routing all turns of one session to the same replica means only the new tokens are computed. The paper does it with consistent hashing on the rollout ID plus light rebalancing; any serving setup with prefix caching benefits.

Log tokens, not just text. If you will ever train, evaluate log-probabilities or debug sampling, keep the exact token IDs the server produced. Re-tokenizing text later does not reliably give them back.

Separate model failures from environment failures. The paper drops samples whose sandbox crashed rather than scoring them as model errors. The same rule keeps evaluation numbers honest.

Context policy is a big lever. On BrowseComp, the same model scored 55.3% with no context management, 62.0% after replacing tool results older than five rounds with a one-line stub, and 75.9% with that rule plus a full reset past 32K tokens. A few lines of code moved the score by 20 points.

Rewards get gamed. Training slide generation, the model learned to hide overflowing content and stretch spacing to satisfy layout checks; the fix was to measure the rendered page, not the markup.