Z.ai's GLM-5.3-Flash Pairs Sparse and Linear Attention in an Open-Source First — at $0.15 per Million Tokens

Z.ai's latest open-weight release is less notable for its benchmark scores than for its architecture — though the scores are strong too. GLM-5.3-Flash, released under the MIT license, is the first open-source frontier-class model to combine sparse attention with linear attention in a single hybrid design, a configuration aimed squarely at making million-token contexts economically viable.

The architecture

GLM-5.3-Flash is a 320-billion-parameter mixture-of-experts model that activates just 18 billion parameters per token. Unlike earlier Flash variants that grafted vision onto a text model, it is natively multimodal, trained from a new multimodal base with a 1M-token context window spanning text and images.

The hybrid attention scheme is the headline contribution: at long context, Z.ai reports roughly 3x lower attention compute and over 4x smaller KV cache versus the base GLM-5.3 — the two costs that dominate serving economics for agentic and long-document workloads.

Benchmarks and price

On Z.ai's DeepSWE software-engineering evaluation, GLM-5.3-Flash scores 63.4 against GLM-5.2's 46.2, with a larger margin on AutomationBench. Self-reported numbers deserve the usual skepticism, but early community testing has been notably positive — the model trended for days after appearing unannounced as a stealth checkpoint on evaluation platforms.

Then there is the price: $0.15 per million input tokens and $0.50 per million output, with cached input at $0.03. Against GPT-6 Astra's $10/$50 and Fable 5.1's identical frontier pricing, Z.ai is offering roughly 90 percent of the capability surface for coding and agentic work at around 1-2 percent of the cost.

The pattern

GLM-5.3-Flash extends the strategy that has defined China's leading labs all year: ship open weights fast, compete on serving efficiency rather than raw scale, and occupy the default slot in cost-sensitive stacks worldwide. With the GLM line already powering coding agents and now a Saudi national platform built on rival MiniMax's weights, the open-weight efficiency race is arguably where Chinese labs are most clearly ahead — and hybrid attention just raised the bar again.