DeepSeek's New Model Cuts Memory Costs by 75% — and It's Free to Download
DeepSeek just shipped another model, and this one is worth paying attention to for a reason that isn't about raw benchmark scores. It's about cost — specifically, the cost of remembering.
DeepSeek-V4.1-Flash is a 552-billion-parameter model (with an extra 196 billion "Engram" parameters bolted on) that handles a 1-million-token context window while using roughly a quarter of the memory its predecessor needed to do the same job. It's open-sourced under an MIT license, and within hours of release it was already live on SiliconFlow, WorkBuddy, and the DeepSeek API simultaneously.
Why "Memory" Is the Real Story Here
Every AI model that holds a long conversation or reads a long document has to store something called a KV cache — a running record of everything it's "read" so far, so it doesn't have to reprocess the whole conversation on every new word. The longer the context window, the bigger that cache gets, and the more expensive (and slower) the model becomes to run.
This has been the quiet bottleneck behind every "give the AI a bigger context window" announcement for the past two years. A bigger window is easy to promise. Making it affordable to actually use is the hard part.
V4.1-Flash attacks that problem directly. Its global KV cache footprint is about 890 bytes per token — roughly a quarter of what DeepSeek-V4-Flash needed, and by DeepSeek's own math, about 1/437th of what the original DeepSeek-V1 needed. That's not a rounding error. That's the difference between a 1-million-token conversation being a novelty and being something you can actually afford to run in production.
What's Under the Hood
A few technical choices make this possible:
- A new Causal Encoder-Decoder architecture — a structural change from prior DeepSeek releases, built specifically around efficient long-context handling.
- FP4 KV caching and cross-layer attention reuse — instead of storing a full-precision copy of attention data at every layer, the model shares and compresses it, cutting memory without (DeepSeek claims) a meaningful accuracy hit.
- Native multimodal vision support — it's not text-only; it processes images natively as part of the same architecture.
- Sparse activation — despite being 552 billion parameters on paper, only about 8 billion activate during prompt processing and 16 billion during generation, which is what keeps inference costs down despite the model's size.
On benchmarks, DeepSeek reports a GPQA Diamond score of 90.9, an HLE score of 36.8, a Codeforces rating of 3471, and a 90.6 on Terminal-Bench 2.1 — the last one being a proxy for how well a model performs at actual coding-agent tasks, not just answering trivia.
Free, Fast, and Already Everywhere
Two things make this release different from a typical model drop. First, it's MIT-licensed — no usage restrictions, free to self-host, free to build on. Second, the rollout was immediate: SiliconFlow had it live on day zero, WorkBuddy is offering two weeks free, and DeepSeek's own API dropped prices alongside the release.
That combination — open weights, cheap inference, and same-day availability across multiple platforms — is becoming DeepSeek's signature move. It's less about topping a leaderboard for a week and more about making a genuinely capable model cheap enough that everyone downstream can build on it immediately.
What This Means If You Use OpenClaw
This kind of efficiency work matters more than it looks like at first glance, especially for anyone running AI agents rather than single chat exchanges.
An agent that works through a long task — reading documentation, running commands, checking results, adjusting its plan — is constantly accumulating context. Every tool call, every file it reads, every step it takes adds to what it has to keep track of. A model that can hold a million tokens of that history affordably is a model that can stay useful through a genuinely long, multi-step job instead of losing the thread or needing the conversation truncated halfway through.
OpenClaw runs on exactly this principle: agents that maintain context across real, extended work — not just a single reply, but a whole session of tool use, memory, and follow-through. Cheaper long-context models are part of what makes that kind of sustained, autonomous work more practical for everyone, not just teams with big infrastructure budgets.
If you want to see what an agent that actually holds onto context through a real task looks like, you can try it yourself.