- Z.ai attributes GLM-5.3's improvements over 5.2 to post-training on the same base model.
- The model's public benchmark table includes both gains and comparisons in which rivals remain ahead.
- Weights are now available, and Mistral added an unmodified hosted version in September.
Post-training drives the new version
Z.ai's GLM-5.3 launch emphasizes coding and work that requires many successive actions. The company says it uses the same base model as GLM-5.2, with the improvements coming from post-training. Its repository lists the model as 744B-A40B, meaning 744 billion total parameters and 40 billion activated parameters.
The headline 50% coding improvement refers to Z.ai's internal Code Bench. It should be attributed to the company and to that particular evaluation. It is not a claim that every programming task is completed 50% faster or with 50% fewer errors.
The public results offer a more detailed picture
The released model card reports 88.2 on Terminal-Bench 2.1, compared with 81.0 for GLM-5.2, and 66.9 on DeepSWE versus 46.2. On NL2Repo it lists 58.0, the same as Kimi K3 in that table and below DeepSeek V4 Pro's 61.1. The comparisons show a stronger model without supporting an unqualified lead on every coding measure.
The card also documents how runs were conducted. Terminal-Bench 3.0 uses the Claude Code harness, maximum reasoning effort, repeated rollouts and a ten-hour timeout. That is a substantial execution budget. A score from such a run describes the model working through an agent environment over time, rather than its first answer to a short prompt.
Access has expanded since the launch
Weights that were initially announced for a later release are now downloadable through Hugging Face and ModelScope. The repository lists FP8 and BF16 variants, as well as local-serving routes. The published model card carries a GLM-5.3-specific license; it should not be assumed to have the same license as GLM-5.3-Flash.
A further distribution route arrived through Mistral. Its documentation dates the hosted GLM-5.3 offering to September 15 and labels it a public preview of a third-party model served without Mistral modifications. The service offers a one-million-token context and up to 128,000 output tokens.
GLM-5.3 and Flash are different deployment choices
The hosted GLM-5.3 listing at Mistral describes a text model. That is distinct from the natively multimodal GLM-5.3-Flash release, whose developer documentation includes image and video inputs. The version number alone is therefore not enough to choose an endpoint for a workflow that needs to inspect screenshots or footage. A customer should compare the actual model card and host's supported features.
For local GLM-5.3 serving, Z.ai also lists Ascend NPU routes alongside GPU-oriented frameworks. Its model card documents low, high and max reasoning budgets, with max used by default and for benchmark reproduction. This setting affects how much computation the model can spend before returning a result. Reducing the budget may suit an interactive product, but the resulting experience should not be presented as a reproduction of a maximum-effort score.
A separate service, with separate terms
Mistral currently lists $1.40 per million input tokens, $0.14 for cached input and $4.40 for output, and supports function calling, structured output and batching. Those are Mistral's hosted terms, not a universal GLM-5.3 price.
The resulting access choices are materially different. Downloaded weights require the operator to provide serving infrastructure; a hosted endpoint supplies that infrastructure under the provider's limits and billing. Neither arrangement includes an entire autonomous work application by itself. The release expands the model available to those systems, while its benchmark documentation shows how much the surrounding tools and execution budget contribute to the reported outcome.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01Updates & corrections
— Expanded benchmark methodology and updated the launch report to reflect subsequent model-weight availability and Mistral hosting.
