A different division of labour.
- DeepSeek reports 8B active parameters for prefill and 16B for decode.
- The decoder’s global KV is projected from final encoder states.
prompt → encoder → shared global KV → decoderLANGUAGE / DEEPSEEK-V4.1-FLASHA vast context.
A vast context.
A smaller footprint.
DeepSeek’s V4.1 Flash combines a causal encoder–decoder with compressed sparse attention and Engram memory. Its design separates how context is built, retained and read during generation.
DEEPSEEK-V4.1-FLASH↗
OPEN THE 3D ARCHITECTURE ATLAS BACKBONE552B parameters
ENGRAM196B parameters
ACTIVE PARAMETERS8B prefill / 16B decode
GLOBAL KV890 bytes / token
INSIDE THE MODEL / 3D ARCHITECTURE ATLASMore context.
More context.
Less repetition.
Build the prompt once. Share its global memory. Spend the next layer on the next decision.
Drag to rotate · Arrow keys supported
MODULE 01 / CAUSAL ENCODER–DECODER
20 + 20
ENCODER + DECODER LAYERSA different division of labour.
A 20-layer causal encoder feeds a 20-layer decoder. Most prompt tokens take the encoder path; the final 128 prompt tokens are replayed to approximate the decoder’s local attention state.
prompt → encoder → shared global KV → decoderREADING THE DIAGRAM
Grouped schematic of the causal encoder–decoder. Encoder states supply decoder global KV; local attention and Engram follow separate paths. The animation is explanatory; no model inference is executed.
BACKBONE MAP
40 layers. A deliberate pattern.
2 Local SWA4 Full4 Reindex30 Reuse
Encoder: layers 1–20. Decoder: layers 21–40. Full, Reindex and Reuse refer to global attention; local SWA is also present.
A CLOSER LOOK / DEEPSEEKThe details behind
The details behind
the diagram.
Configuration, memory and computation—each with a specific role in the model.
Full. Reindex. Reuse.
- Main sparse selection: 512 entries, alongside local SWA.
- Later decoder indexers use at most 16,384 candidate positions.
shared KV → candidate pool → top 512 → attentionComputation and memory, separated.
- 552B backbone parameters and 196B Engram parameters are reported separately.
- Engram is learned model memory, not persistent user memory.
token → top-6 experts + shared / gated n-gram lookupBring visual context into the stream.
- 14-pixel patches and two-dimensional rotary positions.
- The released model accepts image and text inputs and generates text.
image → ViT → pixel unshuffle → projector → backboneFOLLOW THE RESEARCHGo straight
Go straight
to the source.
Developer papers, released configurations and reference implementations. Reviewed 10 September 2026.
D1
TECHNICAL REPORT
D2Pushing the Limits of KV Cache Compression
§§2.1–2.4, 3.2, 4.2.1 · pp.7–14, 19–22
CHECKPOINT
D3Official V4.1 Flash model card & config
Exact V4.1 release; linked configuration and reference code
SOURCE CODE
D4DeepSeek reference inference implementation
CSA2 cache ownership, index reuse, Gate and Engram
RELEASE NOTE
DeepSeek release changelog
September 10, 2026 · DeepSeek-V4.1-Flash
WHAT THIS GUIDE COVERS
Only later decoder indexer scoring is bounded by the candidate cap; the complete model is not constant-cost with context.
SWA replay is approximate. Architecture figures are DeepSeek’s published specifications, not Perchy benchmarks.
MADE MEANINGFUL IN YOUR BUSINESSFrom capability
From capability
to possibility.
Long-context reasoning
Work across substantial documents and multi-step problems with inference capacity sized for the actual context.
Coding and agents
Connect model reasoning to approved tools and carefully controlled production workflows.
Multimodal knowledge
Explore applications that combine language and visual information within the model’s supported input scope.
Know your inference service scope.
Perchy provides managed AI inference through authenticated APIs. Customers do not receive SSH, root, operating-system or bare-metal access. The service does not provide model-training or fine-tuning functionality. Server administration remains with Perchy.
The technical report’s backbone and Engram parameter figures are shown separately. Reported global KV storage is not total GPU memory: local windows, activations, weights, runtime buffers and concurrency add costs. API availability and the licensed serving configuration are agreed in your proposal.
Read the model developer’s documentationKEEP EXPLORINGAnother model.
All models Another model.
Another way to compute.
One conversation.
A whole new trajectory.