DeepSeek-V4.1-Flash Ships With 552B Parameters and 890-Byte KV Cache

TL;DR

• DeepSeek-AI dropped DeepSeek-V4.1-Flash on Hugging Face, a 552 billion parameter open Mixture-of-Experts model, licensed under MIT.
• It handles context windows up to 1 million tokens but only activates 8B parameters during prefill and 16B during decode.
• A new Causal Encoder-Decoder architecture with Compressed Sparse Attention 2 shrinks the global KV cache to 890 bytes per token, roughly a fourfold cut compared to DeepSeek-V4-Flash.
• The model trained on 45 trillion tokens and is free for anyone to download, fine-tune, or deploy commercially.

A 552 Billion Parameter Model That Runs Like a Fraction of Its Size

DeepSeek-AI published DeepSeek-V4.1-Flash on Hugging Face on September 10, 2026, and the repository reads like a dare to the rest of the industry. The model carries 552 billion backbone parameters, wrapped in a Mixture-of-Experts design, released free under the MIT License for anyone to download, fine-tune, or run in production. That’s not a typo. A model this large is usually the kind of thing you rent by the hour from a cloud provider, not the kind you clone to a laptop over a weekend.

The headline spec is the context window: up to 1 million tokens. But the more interesting number sits underneath it. DeepSeek-V4.1-Flash only activates 8 billion parameters during prefill and 16 billion during decode, thanks to what the company calls a Causal Encoder-Decoder architecture. The model card puts it plainly: “We introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens.”

The trick is a technique the team calls Compressed Sparse Attention 2, which projects the decoder’s key-value cache straight from the encoder’s states instead of storing a full copy for every token. The result is a global KV cache footprint of just 890 bytes per token. Compare that to how most large context models bloat their memory usage as the conversation gets longer, and the number starts to look less like an optimization and more like a different way of building the thing entirely. The model trained on 45 trillion tokens before any of this shipped.

Why 890 Bytes Might Matter More Than 552 Billion Parameters

Parameter counts make headlines. Memory footprints make budgets. DeepSeek-AI’s own numbers claim a roughly fourfold reduction in global KV cache size compared to its predecessor, DeepSeek-V4-Flash, while the company says it holds onto competitive agentic and reasoning performance against frontier proprietary models. If that claim survives independent testing, it changes the math for anyone who has ever watched their inference bill balloon the moment a context window stretched past a few hundred thousand tokens.

Here’s a way to picture it. Long-context inference has always worked a bit like a courthouse stenographer who has to keep a full paper transcript of every word spoken in a trial that runs a million pages. Most models file that transcript in a warehouse of cabinets, one per page, and the warehouse gets bigger every time someone talks. DeepSeek’s approach is closer to handing that same stenographer a single notepad that holds the whole trial, because the notepad only records what actually needs recalling later, not every syllable that got spoken. Big claim. Bigger receipts.

I’ve read a lot of model cards this year that lead with parameter counts because parameter counts are easy to market. This one buries its most consequential number, 890 bytes per token, three paragraphs down, and that tells me DeepSeek-AI is building for engineers who already know what KV cache bloat costs them, not for a press cycle.

Does a smaller active parameter count during inference mean a worse model? That’s the question every skeptic will ask, and DeepSeek-AI’s answer is architectural rather than rhetorical: separate the encoder’s job from the decoder’s job, let the decoder borrow compressed context instead of rebuilding it, and only wake up 8 billion or 16 billion parameters at a time instead of all 552 billion. Whether that holds up under third-party benchmarking outside the company’s own claims is the thing to actually watch, but the engineering logic is coherent enough to take seriously rather than dismiss as another benchmark-chasing release.

DeepSeek’s Pattern: Efficiency First, Marketing Second

DeepSeek-AI has built its reputation on squeezing more capability out of less compute, and this release fits that pattern rather than breaking from it. The company’s open-weights models have consistently chased architectural efficiency and reduced inference overhead instead of simply scaling parameter counts and hoping the bill works itself out later. DeepSeek-V4.1-Flash is the latest entry in that line, following DeepSeek-V4-Flash, and the model card frames the new release explicitly as a successor built to cut that predecessor’s KV cache footprint down further.

What makes this notable is the license. MIT isn’t a research-only license with fine print buried in an appendix. It’s about as permissive as open-source terms get, and it means a startup, a hobbyist, or a competing lab can take these weights and build a product on top of them without asking anyone’s permission first. Is that a genuine gift to developers, or just the cheapest way to seed adoption before the next model drops? For an industry where the biggest players still guard their best models behind APIs and usage tiers, that choice alone carries weight, regardless of how the benchmarks eventually shake out.

What Happens Once Outsiders Get Their Hands On It

The real test starts now, once developers outside DeepSeek-AI start running DeepSeek-V4.1-Flash against their own workloads instead of the company’s model card. Watch for independent benchmarks on the million-token context claim specifically, since long-context performance tends to degrade in ways a headline number rarely captures, and the gap between “supports a million tokens” and “reasons coherently across a million tokens” is where a lot of models quietly fall apart. Watch too for how quickly the major inference platforms add support for the Causal Encoder-Decoder architecture, since a novel design like Compressed Sparse Attention 2 only pays off in practice once the serving frameworks people actually use know how to exploit it.

And keep an eye on how proprietary labs respond. If a 552 billion parameter open model really does hold competitive reasoning performance while running this much leaner on memory, that’s a pricing argument every closed-model provider will have to answer, whether they want to or not.

Source: Hugging Face