Z.ai Open-Sources GLM-5.3; Model Found 2,436 Bugs Across 269 Projects
GLM-5.3's open-sourcing puts a frontier Chinese model built on domestic accelerators into public hands, with demonstrated ability to find thousands of real software bugs. The Flash variant achieves near-comparable agentic performance at a fraction of the cost of frontier alternatives.
Z.ai released GLM-5.3's weights publicly, with the model available for download and runnable on Tinker; it identified 2,436 bugs across 269 open-source projects. GLM-5.3 was released August 14, 2026, sharing the same base architecture as GLM-5.2 with all capability gains from post-training. On KingBench 3, an independent fixed benchmark covering coding, simulation, math, and generative tasks, GLM-5.3 scored 91.25%, outperforming Opus 5 and Kimi K3 while performing roughly on par with Fable 5 on tough 3D and UI tasks. Z.ai's benchmark data shows GLM-5.3 outperformed models from Anthropic and OpenAI on a cyber performance test. Together AI ran 904 DeepSWE rollouts: against Fable 5, GLM-5.3 tied on pass@1 and won on pass@4, costing $3.99 per rollout versus $21.63; against GPT-5.6 Sol, it trailed by 3.7 percentage points on pass@1 but won on pass@4 at roughly half the cost, with a GLM-first cascade hitting 85.9%.
Z.ai's Ox Alpha preview was GLM-5.3-Flash, a 320B-parameter mixture-of-experts model with 18B active parameters running on domestic Chinese accelerators. It is MIT-licensed, natively multimodal, and supports a 1M-token context window; it is available on CoreWeave's Serverless Inference platform and on Databricks. On Databricks' OfficeQA Pro v2 benchmark, GLM-5.3-Flash delivers 10% higher quality than GLM-5.2 at one-tenth the cost, with multimodal support for reading figures and verifying web pages in coding contexts. Architecturally, it uses 45 layers versus GLM-4.5's 92, combining linear attention, sparse attention, and a lightweight indexer that cuts attention compute 3x and per-layer KV cache 4.4x versus GLM-5.3. Together AI's 900-rollout comparison shows Flash costs 17x less than the full model, giving up 5.6 pass@1 points, narrowing to 2.6 at pass@4.
Z.ai released GLM-5.3's weights publicly on August 30, completing the open-source step the company had indicated would follow safety evaluation.
The model, released August 14 with capability gains from post-training over a shared GLM-5.2 base architecture, identified 2,436 bugs across 269 open-source projects; weights are available for download and the model runs on Tinker.
Three independent comparisons published August 28 put Z.ai's GLM-5.3-Flash at roughly 17x less cost than the full GLM-5.3, with the pass@1 gap narrowing from 5.6 to 2.6 points at pass@4 in a 900-rollout DeepSWE test and 16.7x less on a Blender MCP task.
A three-model price test placed Flash at $0.03 versus $0.29 for Claude Opus 4.6 on the same prompt. Rohan Paul found that wrapping full GLM-5.3 in the Atomic Agent harness roughly doubled token usage while adding only $0.77 to cost.
GLM-5.3-Flash became available on Databricks on August 27, adding a second major cloud platform alongside CoreWeave for Z.ai's 320B-parameter mixture-of-experts model.
Databricks reports it delivers 10% higher quality than GLM-5.2 at one-tenth the cost on OfficeQA Pro v2 and adds multimodal support for reading figures and verifying web pages. A Hugging Face analysis the same day found Chinese labs are releasing open-weight models several times larger than any open models US labs have shipped publicly in 2025-2026.
Together AI published 904-rollout DeepSWE comparisons August 27 showing Z.ai's GLM-5.3, released August 14, tied Fable 5 on pass@1 and beat it on pass@4 at 5.4x lower cost; KingBench 3 placed it at 91.25%, ahead of Opus 5 and Kimi K3.
Bloomberg Intelligence analyst Robert Lea said Z.ai 'remains on a completely unsustainable commercial footing,' with shares down 9% on the launch day, and founder Tang Jie argued post-training is now the highest-value scaling dimension.
Together AI: GLM-5.3 trailed GPT-5.6 Sol on pass@1 by 3.7 points, won pass@4 at half the cost; GLM-first cascade reached 85.9%
Tang Jie argued parameter count is an incomplete capability metric and positioned GLM-5.3 as a post-training scaling experiment
Hugging Face analysis finds Chinese labs releasing open-weight models several times larger than any US labs have shipped publicly in 2025-2026
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free