DeepSeek Ships a 552-Billion-Parameter Model With No VRAM Numbers

DeepSeek released DeepSeek-V4.1-Flash on Sept. 10, a 552-billion-parameter model licensed under MIT terms that activates as few as 8 billion parameters for each token it processes, according to the model's Hugging Face repository.

An encoder-decoder split most open models skip

The model uses a mixture-of-experts design, which routes each token to a subset of the model's parameters instead of all of them: one shared expert plus 384 routed experts per layer, with six routed experts activating per token, the model card says. During text generation, 16 billion parameters activate per token, up from 8 billion during the initial prompt read. DeepSeek built the model on what it calls a Causal Encoder-Decoder architecture, a 40-layer transformer split into a 20-layer encoder and a 20-layer decoder, a departure from the decoder-only design of most open-weight language models.

DeepSeek trained the model on a 45-trillion-token multimodal corpus, then extended its context window from an initial 64,000 tokens to 1 million tokens over a further 34 trillion tokens of training, according to the same documentation. Post-training followed what DeepSeek describes as a standard supervised fine-tuning, reinforcement learning and on-policy distillation sequence. A second-generation attention scheme DeepSeek calls Compressed Sparse Attention cuts the key-value cache, the memory a model uses to track earlier tokens in a conversation, to about 890 bytes per token, roughly a fourth of the footprint DeepSeek reports for the prior DeepSeek-V4-Flash release.

The model processes images natively through a vision encoder DeepSeek calls DeepSeek-ViT, trained from scratch rather than adapted from an existing computer-vision model, the card says. It also exposes a "continuously controllable reasoning effort" setting on a scale of 1 to 100, letting a caller trade latency against how long the model reasons before answering, though the documentation does not say how the two ends of that scale compare in tokens spent or wall-clock time.

What the model card leaves out

DeepSeek's README lists no VRAM requirement, no inference-speed figure, no weight-conversion command and no hosted pricing, HackerNoon reported after examining the release on Sept. 14. The model also has no standard Transformers chat template; the repository ships a custom encoding script that developers have to run separately to format prompts, according to files listed in the model's own repository.

Measured against DeepSeek's own larger model, the 1.6-trillion-parameter DeepSeek-V4-Pro-Base, V4.1-Flash trades peak reasoning quality for that efficiency: HackerNoon's review of the released benchmark numbers found it scores lower on the MATH and LongBench-V2 tests, though it beats the older DeepSeek-V4-Flash on coding tasks and multimodal handling, according to the same review.

Model	Total parameters	Active per token (decode)
DeepSeek-V4.1-Flash	552B	16B
DeepSeek-V4-Pro-Base	1.6T	Not disclosed

The MIT license puts the weights themselves within reach of anyone who can host them, unlike the closed-weight GPT-6 Sol and Luna models OpenAI priced this week, or Claude Opus 5.5, which Anthropic keeps proprietary. DeepSeek has not said when, or whether, it will publish the VRAM and latency figures needed to estimate what actually running V4.1-Flash costs.