GLM
Z.ai
Track the attention changes separately from the post-training gains.
MiniMax M3 / Lab profile
MiniMax
Study how sparse attention changes what the GPU reads.
First time comparing models? Learn the terms before the architecture →
M3 is the MiniMax checkpoint covered in this edition. Its native multimodal training, million-position configuration and sparse-attention design make it a useful model for reasoning about long agent tasks. The June announcement and later open-weight availability are separate events.
Architecture trajectory
What connects these models?
MiniMax Sparse Attention uses a lightweight indexer to select relevant KV blocks. The key distinction for a learner is between retained context and the subset read for a query. Sparse reads can reduce attention bandwidth without shrinking every part of the stored cache by the same factor.
Store the history. Read the selected parts.
Fewer bytes read can reduce bandwidth pressure.
That does not prove an equal reduction in stored bytes.
Newest first / grouped sequence, not a time scale
The MiniMax M3 timeline
MiniMax-M3
Sparse reads do not automatically mean a small cache
The June model combines 428B total parameters, roughly 23B active and native multimodal training. MiniMax Sparse Attention adds a lightweight indexer to select blocks; the full retained state still needs its own capacity budget.
The announcement promised weights within ten days. We verified the current open checkpoint, not a precise first-publication day.
Which other labs should you compare?
GLM-5.2 is useful for comparing index sharing with block selection. Kimi K2.6 supplies a contemporary multimodal MoE alternative. Keep prompt lengths and the agent harness fixed before attributing a difference to the attention mechanism.
Kimi
Moonshot AI
Follow the jump from a coding update to a new attention stack.
Interview preparation
What you should be able to explain
Draw a large retained cache, a small block index and the selected reads. Explain why reducing read traffic and reducing allocated memory are distinct achievements.
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.