GLM 5.1

Z.ai's previous flagship for agentic engineering, with significantly stronger coding than its predecessor. It is designed for long-horizon autonomous work, sustaining optimisation over hundreds of rounds and thousands of tool calls rather than plateauing early, with performance that improves as the task lengthens -- Z.ai cites sustained execution for up to eight hours. Thinking is on by default and can be disabled.

The framing of the 5.1 post is a claim about time rather than about peak scores. Z.ai argues that earlier models, its own GLM-5 included, exhaust their repertoire early: they apply the familiar techniques, take a quick gain, then plateau, and giving them longer does not help. GLM-5.1 was built so that extra runtime keeps paying, and the post spends most of its length demonstrating that rather than listing results.

It uses three tasks with progressively less structured feedback. The first is an open-source vector-database challenge scored purely on queries per second at 95% recall, normally run inside a 50-turn tool budget where the best result on record was about 3,500 queries per second. Z.ai restructured it as an outer optimisation loop and reports GLM-5.1 still finding real improvements past 600 iterations and 6,000 tool calls, finishing at 21.5 thousand queries per second. The trajectory is a staircase: six structural rewrites, each chosen by the model after reading its own benchmark logs, with recall temporarily broken around each transition and then restored.

The second is more honest about the ceiling. On the hardest level of a GPU kernel benchmark, where whole architectures have to be optimised end to end, Z.ai reports GLM-5.1 reaching a 3.6 times geometric-mean speedup and still improving late in the run, while naming Claude Opus 4.6 as the stronger model in that setting at 4.2 times with headroom left. The third has no metric at all: an eight-hour run building a Linux-style desktop as a web application, wrapped in a harness that asks the model after each round what is still missing.

The benchmark numbers behind that: Z.ai reports 58.4 on SWE-bench Pro, 42.7 on natural-language-to-repository generation and 69.0 on Terminal-Bench 2.0 with its best harness, against GLM-5's 55.1, 35.9 and 56.2, plus 68.7 on the CyberGym vulnerability-reproduction set against GLM-5's 48.3.

The post ends on what is not solved, which is the part worth reading before you plan a long autonomous run on it: escaping a local optimum earlier once incremental tuning stops paying, holding coherence across execution traces of thousands of tool calls, and, the one Z.ai flags as most important, reliable self-evaluation where there is no number to optimise against. It calls this a first step in that direction. The weights are MIT-licensed and can be self-hosted; the copy served here is Z.ai's.