GLM-5.3
An open-source large language model released by Zhipu AI, sharing the same base as GLM-5.2 but achieving significant improvements in programming capabilities through post-training. It supports local deployment via GGUF quantization format.
Core Facts
Timeline (last 90 days)
GLM-5.3's weights are open-source, and 32 providers offer the model on OpenRouter.
GLM-5.3 has an input multiplier of 6.9, a cached input multiplier of 1.7, and an output multiplier of 24; GLM-5.3-Flash has an input multiplier of 2.3, a cached input multiplier of 0.56, and an output multiplier of 8.
Z.ai charges $1.40 per million tokens for fresh input to GLM-5.3 and $0.26 per million tokens for cached input, a difference of about 5.4 times.
GLM-5.3's thinking tokens count as output tokens and are billed at the 24 multiplier, while the model uses its highest reasoning effort by default.
Anthropic previously released a report stating that the open-source model GLM-5.3 poses a major cyberattack risk.
On Terminal Bench v2.1, Beam's 80.1 trails DeepSeek V4.1 Flash's 90.6; on DeepSWE v1.1, Beam's 44.4 trails GLM-5.3's 61.0 and Kimi K3's 68.0.
在 Surge AI 组织的盲评中，专业标注员给 Mistral Large 4 预览版打出 3.74/5，在五个模型中排名第二，落后于 Claude Opus 5（4.22），领先于 GLM-5.3、Kimi K3 和 GLM-5.2。
Some models enable reasoning by default and count reasoning tokens toward max_tokens, meaning max_tokens is the total budget rather than a cap on the response.
When setting max_tokens=64 for glm-5.3, completion_tokens uses all 64, with 60 spent on reasoning and only 4 tokens left for the response, resulting in empty text.
On the CyberGym benchmark, Mistral Large 4 ranked first among all open-source models with a score of 82, ahead of MIMO, Grok, GLM 5.3, Kimi K3, and others.
40 more timeline events
All Facts (20)
vLLM provides Day-0 support for GLM-5.3, enabling developers to deploy the model on the day of its release.
75%VerifiedZhipu released GLM-5.3, claiming its coding capability improved by 50% compared to GLM-5.2.
65%UnverifiedThe actionable method for validating a coding model is to select a real repair task with tests, fix the repository version and tool permissions, have candidate models complete the same task, and compare first-pass rate, patch readability, per-run cost, and number of human takeovers.
70%UnverifiedGLM-5.3 and GLM-5.2 use the same base, architecture, total parameters, and activated parameters, with the only variable being one month of long-horizon context extension and reinforcement learning.
70%UnverifiedGLM-5.3 scored 84.5 on the CyberGym security capability benchmark.
60%UnverifiedGLM-5.3 has an input multiplier of 6.9, a cached input multiplier of 1.7, and an output multiplier of 24; GLM-5.3-Flash has an input multiplier of 2.3, a cached input multiplier of 0.56, and an output multiplier of 8.
50%UnverifiedGLM-5.3's weights are open-source, and 32 providers offer the model on OpenRouter.
50%UnverifiedZ.ai charges $1.40 per million tokens for fresh input to GLM-5.3 and $0.26 per million tokens for cached input, a difference of about 5.4 times.
50%UnverifiedGLM-5.3's thinking tokens count as output tokens and are billed at the 24 multiplier, while the model uses its highest reasoning effort by default.
50%UnverifiedAnthropic previously released a report stating that the open-source model GLM-5.3 poses a major cyberattack risk.
50%UnverifiedOn Terminal Bench v2.1, Beam's 80.1 trails DeepSeek V4.1 Flash's 90.6; on DeepSWE v1.1, Beam's 44.4 trails GLM-5.3's 61.0 and Kimi K3's 68.0.
50%Unverified在 Surge AI 组织的盲评中，专业标注员给 Mistral Large 4 预览版打出 3.74/5，在五个模型中排名第二，落后于 Claude Opus 5（4.22），领先于 GLM-5.3、Kimi K3 和 GLM-5.2。
50%UnverifiedWhen setting max_tokens=64 for glm-5.3, completion_tokens uses all 64, with 60 spent on reasoning and only 4 tokens left for the response, resulting in empty text.
50%UnverifiedSome models enable reasoning by default and count reasoning tokens toward max_tokens, meaning max_tokens is the total budget rather than a cap on the response.
50%UnverifiedOn the CyberGym benchmark, Mistral Large 4 ranked first among all open-source models with a score of 82, ahead of MIMO, Grok, GLM 5.3, Kimi K3, and others.
50%UnverifiedGLM-5.3 performs best when designing multi-page webpages with interactive effects, but it carries a 'subscription priority' suffix, resulting in longer queues for non-members.
50%UnverifiedThe attacker used ARTEX in conjunction with multiple large language models including DeepSeek v4.1-Flash, GLM-5.3, Grok 4.6, and Claude Code.
50%Unverified469 tok/s means that generating 500 tokens takes just over 1 second.
50%UnverifiedOn 8 AMD Instinct MI355X accelerator cards, using the vLLM inference framework, the GLM-5.3 model achieved a single-user decoding speed of 469 tok/s.
50%UnverifiedThe report states that the safety protection mechanisms built into GLM-5.3 can be gradually stripped away within a matter of days.
50%