DeepSeek's V4.1-Flash Activates Just 8 of Its 552 Billion Parameters

A 552-billion-parameter backbone, MIT licensed

DeepSeek AI's DeepSeek-V4.1-Flash model card on Hugging Face lists 552 billion backbone parameters, released under an MIT license, with support for context windows up to 1 million tokens. The architecture is a Causal Encoder-Decoder: a 40-layer transformer split into a 20-layer causal encoder feeding a 20-layer decoder, with 1 shared expert and 384 routed experts per mixture-of-experts layer. Six routed experts activate for every token.

Prefill and decode do not spend the same budget

That routing is why the model needs two active-parameter figures instead of one. DeepSeek's card lists 8 billion parameters activating per token during prefill, rising to 16 billion during decode. Both numbers sit far below the 552-billion backbone, which is the entire point of a mixture-of-experts design: the compute bill scales with the small active figure, while the memory bill, because every expert has to sit somewhere accessible even when it is not firing, scales with the full 552 billion.

Run the arithmetic on the backbone alone: 552 billion parameters at two bytes each, the BF16 format DeepSeek's card lists as a supported type, comes to roughly 1.1 terabytes before a single token of context is loaded. No single GPU on the market holds that much memory, which is why a deployment running the model at full precision needs multiple nodes regardless of how few parameters activate per token. The active-parameter figure controls the compute bill; the total-parameter figure controls the hardware bill, and for a model this size, the second number is the one that decides whether a team can run it at all.

A 1-million-token context window makes memory the bottleneck

DeepSeek's card lists FP4 in the E2M1 layout as the format for KV caching, alongside BF16, F32, F8_E4M3 and I8 tensor types for the weights themselves. A cache that grows with every one of a million possible tokens is what makes quantization choices matter more for a model like this than for a short-context chatbot: the weights are large but fixed at load time, while the cache grows with every query length a deployment actually serves.

Independent quantizers have already compressed it, at a cost

By September 2026, Hugging Face user apetersson had published a mixed 2-bit build called DeepSeek-V4.1-Flash-MixedQ2-GGUF, using the IQ2_XXS format for the gate and up projections and Q2_K for the down projections, averaging 2.25 bits per weight across 1,038 tensors. The result comes to 169.92 GB spread over five shards, small enough to load on a single Nvidia B200 (180 GB) or B300 (288 GB), but still too large for one 96 GB RTX PRO 6000 card. The card recommends splitting the model across two RTX PRO 6000 cards instead. It is the same tradeoff ISTA-DASLab made when it squeezed Qwen3.8-27B into 8.4 gigabytes earlier this month, on a far smaller model. The quantization here is not calibrated against real activations either; the card describes using constant importance weighting rather than calibrated activation data, the kind of shortcut that trades some accuracy for the ability to run the model outside a datacenter at all.

GPU	VRAM	Fits the 169.92 GB build?
Single RTX PRO 6000	96 GB	No
Two RTX PRO 6000	192 GB	Yes
Single B200	180 GB	Yes
Single B300	288 GB	Yes

1. Total backbone552 billion parameters
2. Active at prefill8 billion parameters
3. Active at decode16 billion parameters

Source: deepseek-ai/DeepSeek-V4.1-Flash model card, Hugging Face, accessed 2026-09-14

DeepSeek's own benchmark numbers, unverified elsewhere

DeepSeek's model card reports 79.4 on HumanEval pass@1, 93.0 on GSM8K at 8-shot, 74.2 resolved on DeepSWE v1.1, and 90.6 pass@1 on Terminal-Bench 2.1. Every one of those figures is vendor-supplied; no independent party has published its own run of DeepSeek-V4.1-Flash against those same benchmarks. DeepSeek's card also does not state a minimum hardware configuration for running the full, unquantized model, leaving anyone without access to a 180-gigabyte-plus GPU to work that number out from community quantizations instead. Running any of it on non-Nvidia silicon means going through a CUDA compatibility layer Nvidia has not endorsed, not a path DeepSeek documents as supported.

What would change the read

The 552-billion-to-8-billion ratio between total and active parameters is what makes a model this size plausible to run on hardware built for something smaller, but it is also an argument that has to be re-litigated for every mixture-of-experts release: the ratio, the routing overhead, and how many experts a batch of real traffic actually touches at once all move the real memory bandwidth cost away from the headline active-parameter figure. Until someone benchmarks tokens per second on the two-card RTX PRO 6000 configuration DeepSeek's own community has already proposed, the 8-billion-parameter figure describes compute, not the latency a deployment would actually see. DeepSeek shipped the earlier, non-multimodal V4-Flash base model alongside V4.1-Flash, and Nvidia has already published its own compressed build of that base model using its NVFP4 format for Blackwell hardware, a separate data point this piece leaves out because DeepSeek's own card does not confirm the two releases share an identical parameter count.