On September 10, 2026, DeepSeek dropped a model that did something unprecedented in the frontier AI race: the smallest member of a new architecture family outperformed the flagship it was meant to succeed. V4.1 Flash, a 552-billion-parameter MoE model, replaced V4 Pro across the company’s API endpoints within four days — and it did so while costing roughly half as much to serve.
The reason is not a minor optimization or a clever marketing label. It is a full architectural reset. DeepSeek threw out the decoder-only Transformer design it had iterated on through V1, V2, V3, and V4, and replaced it with something called Causal Encoder-Decoder (CED). The result is an asymmetric machine that activates only 8 billion parameters on input and 16 billion on output — a 2x split that reflects how real AI workloads actually behave, rather than how researchers wish they behaved.
The Problem Every Frontier Model Faces
For the past four years, the scaling paradigm has been simple: bigger models, more compute, better benchmarks. But as models crossed the trillion-parameter threshold and context windows stretched into the millions of tokens, two hard truths emerged:
-
Agentic workloads are input-heavy, not output-heavy. When a coding agent ingests an entire repository, multiple tool returns, and a chain of intermediate reasoning steps, the input token count can exceed the output by 100:1 or even 150:1 (as documented in a 2026 University of Michigan / Stanford study). Every token of that input pays the same decoder price as a token of output — a giant waste when 99% of the compute is being spent on “reading” rather than “writing.”
-
KV cache is now the dominant cost. For long-context reasoning traces, the memory footprint of the KV cache exceeds the cost of the model weights themselves. DeepSeek’s own V4 Flash required roughly 3,560 bytes per token of cache — enough to make million-token context economically prohibitive for most developers.
The industry’s answer so far has been incremental: grouped query attention, sliding windows, quantization. DeepSeek’s answer was different: tear the machine in half.
CED in One Diagram
Here is what V4.1 Flash’s 40-layer backbone actually does:
Input (up to 1M tokens)
│
▼
┌─────────────────────┐
│ Encoder (20 layers) │ ← 8B active parameters per token
│ "Read" path only │ Compresses input into summary states
└─────────────────────┘
│
▼
Global KV Cache ← Projected ONCE from encoder output
(890 bytes / token) Shared across all decoder layers
│
▼
┌─────────────────────┐
│ Decoder (20 layers) │ ← 16B active parameters per token
│ "Write" path only │ Generates output autoregressively
└─────────────────────┘
│
▼
Output (up to 384K tokens)
Instead of every layer handling both read and write, the encoder chain exists purely to distill the input into a compact representation, and the decoder chain exists purely to generate. Memory is written once in the encoder and projected into every decoder layer — no more per-layer KV accumulation.
This is not a new idea in machine learning. Encoder-decoder architectures have been standard in translation and speech recognition for a decade. What is new is applying this split to a reasoning-first MoE model with native multimodal understanding, and doing it at a scale where the MoE router itself has learned to specialize experts for distinct reasoning modalities.
CSA2: The Attention Backbone That Makes It Work
CED alone would not be enough. The second half of the efficiency story is Compressed Sparse Attention 2 (CSA2), the attention mechanism that keeps cross-layer memory overhead from erasing the gains of the architectural split.
CSA2 partitions the 20 decoder layers into three functional groups:
| Group | What it does | Memory cost |
|---|---|---|
| Full (layer 1 only) | Stores a complete KV representation per token | Baseline |
| Reindex (intermediate layers) | Extracts salient features from Full and stores only those | ~30% of baseline |
| Reuse (remaining layers) | Directly references Full’s KV without recomputing | ~5% of baseline |
The engineering trade-off is brutal: if every layer used Reuse, the model would produce garbage because shallow and deep layers attend to different parts of the context. But if every layer used Full, the HBM cost would double. CSA2 finds the sweet spot by letting the model learn, during RL training, which layers truly need fresh attention and which can safely inherit.
The result is that global KV cache per token drops from ~3,560 bytes (V4 Flash) to 890 bytes (V4.1 Flash) — a 4x reduction that DeepSeek says translates to 1/4 the HBM and 1/8 the SSD storage compared to the previous generation. For a developer running a 1-million-token reasoning trace, that difference is the gap between a $10 API call and a $2.50 API call.
Why This Matters Beyond DeepSeek
CED is not a trick that only works for one model. It is a general framework for building inference-efficient reasoning models — and DeepSeek has designed it to scale upward. The company’s roadmap, hinted at in the technical report, uses the same encoder-decoder split for a planned V4.1 Pro flagship that will likely push total parameters past one trillion while keeping input-side activation at or below 32B.
For the rest of the industry, the implications are stark:
-
Decoder-only is dead for agentic workloads. OpenAI, Anthropic, and Google have all optimized decoder-only models for the chat paradigm, where input is short and output is long. As agentic coding, research, and analysis workloads become dominant, their cost structure will become increasingly uncompetitive.
-
Inference efficiency = pricing power. DeepSeek’s new off-peak price of $0.15 per million input tokens and $0.60 per million output tokens is roughly one-tenth the price of Claude Sonnet 5 and roughly half the price of GPT-5.6 Luna. That pricing is not a loss leader — it is the direct outcome of an architecture engineered for the actual workloads customers are running.
-
MoE routing is not just for training. V4.1 Flash’s GRPO-trained router specializes experts for mathematical deduction, code synthesis, creative language, and multi-step planning. The same router that makes sparse activation possible also makes the model better at the tasks customers care about. It is a rare example of a design choice that improves both cost and quality simultaneously.
What DeepSeek Is Not Saying
The technical report is silent on one question: what gave DeepSeek the courage to throw out its entire V4 architecture? The answer, based on comments from Liang Wenfeng and leaks from the company’s engineering team, is that DeepSeek Harness (DSH) — the company’s agentic tool-use framework — was outgrowing V4’s capacity. DSH’s RL training loop was producing data that V4 could not process efficiently, and the architecture was the bottleneck.
In other words, the model was not designed and then given a harness. The harness designed the model. That reversal of direction — agent infrastructure first, model architecture second — may be the defining shift of the next two years in AI.
The Bottom Line
V4.1 Flash’s CED architecture is not just a clever engineering trick. It is the first major model release that is designed from the ground up for how AI is actually used in 2026: in long-context agentic workflows where reading dominates writing, and where memory costs dominate compute costs. Whether DeepSeek maintains this lead depends on whether it can scale CED to larger models without losing the cost advantage. But for the first time in years, a Chinese AI company is not just following the frontier — it is defining what the frontier looks like.
