GLM
Z.ai
Track the attention changes separately from the post-training gains.
Kimi / Lab profile
Moonshot AI
Follow the jump from a coding update to a new attention stack.
First time comparing models? Learn the terms before the architecture →
Moonshot’s K2-to-K3 sequence is useful because it contains two different kinds of progress. K2.5, K2.6 and K2.7-Code share the trillion-parameter class while changing the multimodal and coding workload emphasis. K3 changes the underlying deployment problem: a 2.8T expert store, hybrid attention and a newer kernel stack.
Architecture trajectory
What connects these models?
K2’s latent-attention lineage leads into K3’s combination of Kimi Delta Attention, gated latent attention and Attention Residuals. Study the recurrent state and attention cache separately. A prefix-cache implementation must respect both; it cannot treat every layer as a conventional transformer KV array.
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 Kimi timeline
Kimi-K3
A cluster model with a new attention design
K2 used latent attention throughout. K3 combines a recurrent state with selective attention, changing how an engine manages cached prefixes and long requests. Its open release includes supporting kernels and communication work, so adopting the model also means adopting a newer runtime stack.
The July 16 announcement preceded the public weights by eleven days.
Kimi-K2.7-Code
A coding upgrade without a new size class
The release concentrates on coding agents and requires thinking mode in its documented service integration. The tool harness, test feedback and output budget are central to evaluating this generation.
Kimi-K2.6
Keep the hardware fixed; compare the completed work
This is a later K2 multimodal generation, not K3’s hybrid-attention redesign. Matching architecture dimensions makes a fair deployment comparison easier; it does not establish identical weights or quality.
Kimi-K2.5
The multimodal starting point for the K2 series
The released artifact combines quantized experts with other components at different precision. Its 32B active count never meant that the whole model could occupy a 32B-sized memory budget.
Which other labs should you compare?
For coding, compare K2.7-Code with GLM-5.3 under the same agent harness. For the newest large-model architecture, compare K3 with Qwen3.8-2.4T-A95B and account for the latter’s text-only open checkpoint. These are candidate evaluations, not an asserted quality ranking.
Qwen
Alibaba Qwen
One family spans tiny dense models, coding MoEs and conditional memory.
Interview preparation
What you should be able to explain
Trace one agent request through tool feedback, multimodal preprocessing, expert routing and cache reuse. Then explain which costs remain when active compute is much smaller than total storage.
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.