Kimi
Moonshot AI
Follow the jump from a coding update to a new attention stack.
GLM / Lab profile
Z.ai
Track the attention changes separately from the post-training gains.
First time comparing models? Learn the terms before the architecture →
GLM’s 2026 releases show why a version number is not an architecture description. The smaller GLM-4.7-Flash, the 744B-class GLM-5 line and the 320B GLM-5.3-Flash occupy different deployment classes. Within the flagship line, later training and attention changes deserve separate evaluation.
Architecture trajectory
What connects these models?
GLM-5 uses sparse attention at large MoE scale. GLM-5.2 introduces IndexShare across layers; GLM-5.3 builds on that base through post-training. GLM-5.3-Flash instead combines KDA with sparse latent attention and index pooling, adding native visual input in a smaller total-parameter class.
Two kinds of layers retain two kinds of state.
Hybrid attention does not mean zero request state.
A full-attention formula cannot describe every layer.
Newest first / grouped sequence, not a time scale
The GLM timeline
GLM-5.3-Flash
Smaller than GLM-5.3, still a multi-GPU model
The model combines Kimi Delta Attention with sparse latent attention and index pooling. Supporting those operators matters more than whether a server can merely load an FP8 tensor file.
This is the dated Z.ai release-note event; a separate first weight-upload day has not been certified.
GLM-5.3
A post-training release with a changed adoption decision
The card attributes the update to post-training and highlights cyber-defense evaluations. Hosted access and weights did not arrive together; this history matters when reproducing comparisons or dating adoption.
The pinned card dates its launch comparison August 14; the API release notes are dated August 18. Weights were staged later; do not treat August 14 as their upload date.
GLM-5.2
A million-token setting backed by index sharing
Index sharing reuses selection work across layers. It changes the implementation and memory-access pattern, but a million configured positions still require a realistic request budget and a long-context quality test.
This is the dated Z.ai release-note event; a separate first weight-upload day has not been certified.
GLM-5.1
Longer agent runs need process-level evaluation
The release focuses on planning and iterative refinement. The vLLM guide reports the same approximate 744B/40B size class, so deployment scale remains substantial even when post-training changes task behavior.
This is the dated Z.ai release-note event; a separate first weight-upload day has not been certified.
GLM-5
A large sparse-attention foundation for the GLM line
The release emphasizes longer engineering and agent tasks. Later checkpoints refine training and attention behavior, which makes the family a good study of improvement without treating every decimal version as a new size class.
This is the dated Z.ai release-note event; a separate first weight-upload day has not been certified.
GLM-4.7-Flash
A compact way to learn latent-attention serving
The January release occupies a different deployment class from the later 320B GLM-5.3-Flash. “Flash” is a product-family label, not a stable size or latency category.
This is the dated Z.ai release-note event; a separate first weight-upload day has not been certified.
Which other labs should you compare?
Compare 5.2 with 5.3 to study post-training at a similar base scale. Compare Flash with Qwen3.8-Flash-Next for visual agent workloads. For coding, Kimi K2.7-Code is a useful independent candidate, provided the tool harness and reasoning budgets match.
Qwen
Alibaba Qwen
One family spans tiny dense models, coding MoEs and conditional memory.
Interview preparation
What you should be able to explain
Draw the sparse-attention index, selected reads and retained cache as separate objects. Then explain why a model can improve coding results through post-training without reducing serving memory.
Study the underlying systems → · Get the illustrated report →
Sources and scope
Checked 13 September 2026. Assessments and comparison candidates are our interpretation of the cited designs, not measured cross-lab rankings. Each timeline entry links to its release evidence; model pages preserve pinned card and configuration sources.