DeepSeek Releases V4.1-Flash With a New Encoder-Decoder Architecture and MIT-Licensed Weights

On September 10, 2026, the Chinese AI lab DeepSeek released DeepSeek-V4.1-Flash, a 552-billion-parameter mixture-of-experts model that restructures how the model handles input versus output and ships with weights freely downloadable under the MIT license. The release is documented in DeepSeek's official announcement, the model card on Hugging Face, and the company's API changelog.

What changed

V4.1-Flash is built on what DeepSeek calls a Causal Encoder-Decoder (CED) structure: a 40-layer transformer split into a 20-layer causal encoder that processes the prompt and a 20-layer decoder that generates the response, according to the Hugging Face model card. This differs from the single, uniform stack used in most current large language models, where the same layers handle both reading and writing. Splitting the two lets DeepSeek size them differently: of the 552 billion total parameters, only 8 billion activate per token during prompt processing and 16 billion during generation, each routed through 6 of 384 available experts per mixture-of-experts layer.

The practical effect DeepSeek emphasizes is memory. The model card reports key-value cache compression down to 890 bytes per token using an FP4 format for the main cache — about a quarter of what the predecessor, DeepSeek-V4-Flash, required. Lower KV-cache memory per token means a server can hold more concurrent long-context conversations in the same GPU memory, which is the actual bottleneck in production inference, not raw parameter count.

The model supports a context window of up to 1 million tokens, native multimodal input (images alongside text), and a controllable reasoning-effort setting, per the API pricing page. DeepSeek's own benchmark table, published in the September 10 changelog, lists a GPQA Diamond score of 90.9, a Codeforces rating of 3471, and 90.6 on Terminal-Bench 2.1 — figures the company positions as ahead of its own previous flagship, V4-Pro.

Independent verification

Self-reported benchmarks from a model's own maker are not independent evidence, so the more useful check is third-party evaluation. Artificial Analysis, which runs a standardized battery of reasoning, coding, and knowledge tests across commercial and open models, measured V4.1-Flash (in its reasoning, maximum-effort configuration) at 40 on its Intelligence Index, ranking it 6th among 113 models it tracks — well above the sample median of 18. Artificial Analysis also measured throughput at roughly 217 output tokens per second, with the caveat that the model tends to generate long responses: it produced 250 million tokens over the course of the evaluation, against a median of 140 million for comparable models.

Independent measurement matters here because inference cost and speed determine whether a model is actually usable at scale, and those numbers can diverge from a lab's marketing figures once run under uniform conditions.

Pricing and the retirement of V4-Pro

DeepSeek's pricing page lists off-peak output at $0.60 per million tokens and peak output at $1.20, with input tokens priced separately for cache hits ($0.003–$0.006 per million) versus cache misses ($0.15–$0.30 per million). Peak hours are defined as 01:00–04:00 and 06:00–10:00 UTC on weekdays; all other time is billed at half the peak rate. Beginning September 14, 2026, DeepSeek will route all API traffic previously directed to the older V4-Pro model to V4.1-Flash at V4.1-Flash's lower rates, effectively retiring V4-Pro until a V4.1-Pro successor ships.

Why it matters

Three things distinguish this release from a routine version bump. First, the architecture change is a genuine departure, not a scaling exercise: separating the encoder and decoder stacks, and sizing their active parameters differently, is a structural bet that reading a prompt and writing a response are different enough jobs to deserve different amounts of compute. DeepSeek's earlier V4 technical report, published on arXiv, documents the lineage this model builds on, including the lab's prior work on efficient million-token context handling.

Second, the KV-cache reduction is the kind of efficiency gain that changes deployment economics rather than benchmark leaderboards — it lets the same hardware serve more simultaneous long-context sessions, which is what most production AI products actually need.

Third, releasing 552 billion parameters of weights under the MIT license, with no restrictions on commercial use or redistribution, keeps a frontier-class model available for researchers and engineers to inspect, fine-tune, and run outside any single company's API — a contrast to the closed weights of most competing flagship models.

Where to check the numbers yourself

• Official announcement and technical description: deepseek.com
• Full model card and downloadable weights: Hugging Face
• Independent benchmark scores: Artificial Analysis