DeepSeek V4.1 Flash: 552B parameters, 8B active, 60% cached-input cut

DeepSeek V4.1 Flash runs on a new architecture the company calls Causal Encoder-Decoder. Its 552 billion parameters span a 20-layer encoder and 20-layer decoder, but only 8 billion activate on input tokens and 16 billion on output. That makes it structurally leaner than models that activate larger parameter slices for every operation — a design choice that pays off directly in inference cost and memory efficiency.

The memory gains are specific. V4.1 Flash’s KV cache sits at 890 bytes per token, roughly one-quarter of V4-Flash. FP4 caching and Compressed Sparse Attention 2 push the persistent cache footprint to one-eighth of the previous generation. Cached input now costs $0.003 per million tokens at off-peak — down 60% from V4-Flash. Uncached input fell 33%, output 11%.

On the benchmarks developers track most, the model holds its own. DeepSWE v1.1 — the autonomous software engineering test — gives it 74.2, fractionally ahead of Anthropic‘s Opus 5 at 74.0 and above GPT-5.6 Sol at 73.0. GPQA Diamond scores 90.9. The gap that stands out is Terminal-Bench 3.0: 30.0 against Opus 5’s 43.3, a real difference on tasks requiring extended interactive debugging.

Vision is baked in from pre-training. Earlier, V4 split vision into a separate Vision-Exp variant. V4.1 Flash replaces both with a single model built around DeepSeek-ViT, a purpose-trained encoder co-developed with the language model from the start. The multimodal tier is gone — every call gets vision by default.

The context window is 1 million tokens with output capped at 384,000, enough for long-form document analysis and extended agentic workflows without chunking. Existing API callers using the old deepseek-v4-flash and deepseek-v4-flash-vision-exp IDs are already being routed to V4.1 Flash. On September 14, DeepSeek V4 Pro traffic switches over too — at Flash pricing.

DeepSeek described V4.1 Flash as “the smallest model in our new architecture family.” A larger V4.1 Pro on the same Causal Encoder-Decoder design is implied. If it maintains the efficiency trajectory of Flash, it would extend this architecture’s performance-per-dollar further up the benchmark curve — into territory currently held by models that cost significantly more per million tokens.

The architectural point matters beyond the pricing table. KV cache growth is a primary constraint on running large models at long context, and it scales with every token processed. V4.1 Flash’s approach — fewer active parameters, compressed cache — addresses that constraint directly. It is a usable template for long-context deployments, not a corner-case optimization.

Tags: Anthropic, AI models, DeepSeek, LLM, AI pricing, DeepSeek V4.1 Flash