Qwen
Alibaba Qwen
One family spans tiny dense models, coding MoEs and conditional memory.
DeepSeek V4 / V4.1 / Lab profile
DeepSeek
Flash evolves from a post-training update into a different memory system.
First time comparing models? Learn the terms before the architecture →
The April V4 preview, July Flash update, August Pro update and September V4.1 release form a useful chronological study. The July Flash update retains the earlier architecture and size; September changes the architecture. Those events should not be collapsed into one moving Flash label.
Architecture trajectory
What connects these models?
V4.1 Flash uses a causal encoder-decoder, a shared global attention representation and separate Engram memory. Its reported active path differs between prefill and decode. This makes the model a concrete example of why a single parameter count and one all-GPU memory formula can be inadequate.
Prefill and decode do not execute the same path.
Publisher active path: 8B prefill / 16B decode.
552B backbone + 196B Engram is not one flat GPU budget.
Newest first / grouped sequence, not a time scale
The DeepSeek V4 / V4.1 timeline
DeepSeek-V4.1-Flash
The September model that changes the deployment diagram
The card separates 20 causal encoder and 20 decoder layers. It reports 8B active parameters during prefill and 16B during decode, a 552B backbone and 196B Engram memory. Decoder attention reuses a shared global representation produced from encoder output.
DeepSeek-V4-Pro-0813
Keep the production update distinct from the preview
DeepSeek’s release log records the Pro update separately. The inspected checkpoint does not establish its total parameter scope, so the reference preserves the exact identity while withholding unsupported weight arithmetic.
This is the dated Pro service update. The exact public checkpoint is separately pinned below.
DeepSeek-V4-Flash-0731
A post-training update you can compare fairly
The July 31 release updates post-training rather than introducing V4.1’s encoder-decoder and Engram design. Keep the version suffix in evaluation records.
The release note identifies a post-training update with unchanged architecture and size.
DeepSeek-V4-Flash
The smaller V4 preview that made cache design central
The 284B/13B release combines a large sparse expert pool with V4-specific attention and cache behavior. Its one-million-position configuration made cache implementation a first-order deployment concern.
DeepSeek-V4-Pro
Trillion-parameter storage with sparse per-token work
The April preview provides a useful large-versus-small comparison within V4. Later Pro and Flash checkpoints should retain their own identities and evaluation records.
Which other labs should you compare?
Compare Flash-0731 with the original Flash to isolate post-training. Compare V4.1 with Qwen3.8-Flash-Next to study different conditional-memory designs. Kimi K3 is another large hybrid-attention candidate, but no cross-lab quality winner is asserted here.
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 the prefill and decode paths independently. Mark shared model memory, per-request state and data movement. Explain why changing an API alias does not change the identity of an already downloaded checkpoint.
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.