GLM-5.1

A model built around coding and long tasks

GLM-5.1 is a large language model from Zhipu AI (overseas brand Z.ai). Its Hugging Face model card lists 754B parameters in BF16 and F32 tensor types ^[1]. It was released to Zhipu subscribers in late March 2026 and then open-sourced on 7 April 2026 under the MIT License ^[2][1]. MIT is a permissive license that allows reuse, modification and redistribution, so the release put the model's weights in a form others can build on rather than only call through an API.

What Zhipu emphasized

Zhipu's release highlighted two things. First, coding performance significantly stronger than the previous GLM-5. Second, work on long-horizon agent tasks: the model is described as able to sustain optimization over hundreds of rounds and thousands of tool calls ^[1]. The second claim is about endurance rather than a single answer. An agent working on a long task has to keep choosing actions, reading results and correcting course; staying pointed at the same goal across a very large number of steps is a different problem from getting one response right.

Benchmark results

On SWE-Bench Pro, GLM-5.1 scored 58.4, the top entry in its comparison table, ahead of GPT-5.4 at 57.7 and Claude Opus 4.6 at 57.3 ^[1]. Zhipu's model card states that the model achieves state-of-the-art performance on SWE-Bench Pro ^[1]. SWE-Bench-style benchmarks pose real software issues from code repositories and check whether the model can produce a fix that passes the project's tests, so the scores are meant to track practical programming work rather than puzzle-solving.

On Terminal-Bench 2.0 (Terminus-2) it scored 63.5, up from 56.2 for GLM-5, and 69.0 when paired with Claude Code ^[1].

Further gains over GLM-5 were reported on several other tests: NL2Repo at 42.7 versus 35.9, CyberGym at 68.7 versus 48.3, and BrowseComp at 68.0 versus 62.0 ^[1]. HLE rose to 31.0 from 30.5, and HLE with tools rose to 52.3 from 50.4 ^[1]. On Vending Bench 2, which is scored in money, GLM-5.1 recorded $5,634.41 against $4,432.12 for GLM-5 ^[1].

Where it trailed

Not every number was ahead. On GPQA-Diamond the model scored 86.2, behind Gemini 3.1 Pro and GPT-5.4 ^[1]. On HLE with tools its 52.3 was below Claude Opus 4.6 ^[1]. The picture the table gives is a model that is highly competitive on code and agentic execution but not the leader on every reasoning test.

Release and aftermath

Market reaction came the same day it was open-sourced: on 7 April 2026 Zhipu raised API prices by 10%, and its shares rose 11.5% on the release ^[2]. The model had a short run at the top of its line. On 16 June 2026 it was succeeded by GLM-5.2 ^[2], roughly two months after the open-source release.

Where it sits

GLM-5.1 is best read as a step in the fast cadence of open-weight model releases: a widely licensed model aimed squarely at coding tools and agents that run for a long time, shipped with a detailed comparison table and replaced within a couple of months. The headline numbers come from Zhipu's own comparison table and model card rather than independent evaluation ^[1], which is worth keeping in mind when reading the margins over GPT-5.4 and Claude Opus 4.6 — a point or two on a benchmark is a narrow gap. What the release does establish clearly is the direction of the work: better code generation, and agents that can keep working through hundreds of rounds without losing the thread.