Zhipu open-sources GLM-5.1, topping SWE-Bench Pro

Background

GLM-5.1 is a large language model from Zhipu. The model card published on Hugging Face lists 754 billion parameters ^[1]. It is the successor to GLM-5, and Zhipu presented the new version chiefly as an advance in two areas: writing code, and agent work that continues over long stretches of time rather than a single exchange.

What happened

Zhipu first gave its own subscribers access to the model in late March 2026. It then released the weights publicly on 7 April 2026 under the MIT License ^[2][1]. The Hugging Face model card records the same MIT license ^[1]. MIT is a permissive open-source license, which means the model files can be downloaded, used, modified and redistributed with few restrictions, rather than only being reachable through a hosted API.

That two-step pattern — subscribers first, open weights a short while later — is common among labs that want paying users to see a new model before it becomes freely available.

Key details

The benchmark numbers are what Zhipu put at the front of the release. On SWE-Bench Pro, a benchmark built from software-engineering tasks in which a model has to produce a working patch, GLM-5.1 scored 58.4 ^[1]. Zhipu's model card describes this as state-of-the-art performance on that benchmark ^[1]. In the comparison table the company published, 58.4 was the highest figure, ahead of GPT-5.4 at 57.7 and Claude Opus 4.6 at 57.3 ^[1]. The margins are narrow — under a point in both cases — so the table is better read as showing the model in the leading group than as showing a clear separation.

A second set of results covers command-line, terminal-style tasks. On Terminal-Bench 2.0, run with the Terminus-2 harness, GLM-5.1 scored 63.5, up from 56.2 for GLM-5 ^[1]. The same benchmark reached 69.0 when the model was run together with Claude Code ^[1], which suggests that pairing the model with an existing agent scaffold can add several points over the standalone setup.

Beyond the numbers, Zhipu described the model as significantly stronger than GLM-5 at coding, and emphasised long-horizon agent work — sessions that keep optimising a task over hundreds of rounds and thousands of tool calls ^[1]. That description points at the kind of work where a model is expected to plan, call external tools repeatedly, read the results and adjust, rather than answer once and stop.

Impact

The commercial response came the same day. On 7 April 2026 Zhipu raised its API prices by 10%, and its shares rose 11.5% on the release ^[2]. Raising prices alongside an open-weights release is a signal that the company expected demand for the hosted version to hold up even though the weights themselves became free to download.

For people building with the model, the MIT license is the part that matters most: it puts a 754-billion-parameter model ^[1] in a category that can be run, fine-tuned and shipped inside other products without the restrictions of a more limited licence. For the wider field, the release is one more data point in a pattern that has held for coding models — a new version arrives, claims or takes the top of a public benchmark, and the gap to the previous leaders is measured in fractions of a point.