DeepSeek V4.1 Flash: a Causal Encoder-Decoder that activates 8B of 552B on input

DeepSeek published the weights for V4.1 Flash on Hugging Face on September 10, MIT licensed. The headline specs are familiar for the family: 552B MoE, native vision, 1M context. The architecture is not. V4.1 Flash is a Causal Encoder-Decoder, CED, and the number that matters is that it activates about 8B parameters per token during prefill and 16B during decode. For any workload that reads far more than it writes, which is most RAG and most agent loops, that asymmetry is the whole point.

How the CED is laid out

The 40 transformer layers are split in half. The first 20 form a causal encoder that runs over the input. The last 20 form a decoder, and the decoder's global KV cache is not built layer by layer the usual way; it is projected from the encoder's final hidden states. So the expensive full-depth pass happens once over the prompt with a small active parameter set, and generation runs against a compact cache derived from it.

Around that core DeepSeek stacks several efficiency mechanisms:

• FP4 KV cache, which is what makes a 1M-token cache fit anywhere reasonable.
• Cross-layer attention reuse, sharing attention work between layers instead of recomputing it.
• Engram conditional memory, a 196B sparse component that is consulted rather than run.
• DSpark speculative decoding, built in rather than bolted on.

The model was trained from scratch on 45T multimodal tokens; it is not a fine-tune of V4.

Why 8B at prefill changes the bill

Consider a RAG call with 80K tokens of retrieved context and a 300-token answer. In a conventional MoE the prefill runs the full active parameter count over all 80K tokens, and that dominates the cost. In the CED the encoder does that pass with roughly 8B active, and only the 300 output tokens pay the 16B decode rate. Input-heavy workloads therefore see a cost and throughput improvement that output-heavy chat does not. If your agent loop spends its tokens reading tool results and repo files, this is built for you. If it spends them writing long essays, the gain is smaller.

Most agent pipelines are 90% reading and 10% writing, and this is the first frontier-class open model whose architecture prices it that way.

The FP4 cache matters for the same workloads. At 1M context the cache, not the weights, is what runs out of memory first, and halving its footprint again relative to 8-bit shifts the KV cache maths for anyone serving long context on fixed hardware.

Where it sits in the family

The timeline is fast. On September 8 DeepSeek completed the V4 family: V4 Pro went GA, V4 Flash entered public beta, and a V4 Flash-Vision-Exp variant arrived at no surcharge. On September 9 and 10 a V4.1 Flash test endpoint appeared, claiming performance above V4 Pro, and expired on September 10. Then the weights landed. The 2026 DeepSeek lineup now has an open, MIT, input-cheap model at the top of the Flash tier.

One operational detail belongs on your calendar: from September 14 at 12:00 Beijing time the deepseek-v4-pro API endpoint will serve V4.1 Flash at Flash prices until V4.1 Pro ships. Anyone calling v4-pro will get a different model, with a different architecture, under the same name. That is cheaper, and it is also a silent behaviour change for every eval and every prompt tuned against V4 Pro.

What to do this week

• Pin your DeepSeek calls to explicit model IDs before September 14 and re-run your eval set against V4.1 Flash so you have a baseline.
• If you self-host, measure prefill throughput separately from decode; the CED gain lives almost entirely in the first number.
• Test the FP4 cache at your real context lengths; check quality at 200K and 500K, not just at 8K.
• Recompute cost per task for your RAG pipeline with an 8B prefill rate; it likely moves the API-versus-self-host line.

The honest gap

The claim that V4.1 Flash beats V4 Pro came from a test endpoint with no published methodology, and the architecture description is DeepSeek's own. Nobody outside DeepSeek has yet benchmarked the CED against a standard MoE at equal active parameters, or measured what FP4 KV does to retrieval accuracy at 1M tokens. The design is genuinely new; the numbers are still the vendor's until someone runs them.