How much memory does DeepSeek V4.1-Flash need?

The configuration we ship and quote: 196.3 GB total in the Blackwell NVFP4 build — 187.4 GB of weights, a Q8 cache at its native 1M window, and the runtime buffer.

What does dobkra change about DeepSeek V4.1-Flash?

We run the Blackwell-optimised NVFP4 build on NVIDIA hardware. NVFP4 is 4-bit in the tensor cores' native format on 5th-generation cards — the model keeps its quality and gains speed and memory, and it is the build our machines ship with. Quantisation sized from measured model files, not nominal maths — the same rule as every model here. KV cache at Q8_0 by default (half the cache, well under 1% quality cost), which is what keeps the 1M-token window affordable in memory. We run it at its full native 1M window — nothing offered here is a scaled extension. Prompts and tool configurations tuned per role, including Danish documents — the text-heavy roles that never need to see are exactly where this model earns its place. No vision: it is paired with a seeing model for any job that has to read a document, and the blueprints say so rather than leaving it to be discovered.

What context length does DeepSeek V4.1-Flash support?

Native window 1M, extensible to 1M with RoPE/YaRN scaling (a real option with a real quality cost). The cache costs 0.015 GB per 1,000 tokens at full precision, less than half that at Q8.

DeepSeek V4.1-Flash

The value pick: near its bigger sibling's quality for a fraction of the memory, under the most permissive licence in the catalogue, and a native 1M-token window with a cache cheap enough to actually use it. Text and code only — where a job needs to read a document, it runs beside a seeing model rather than instead of one.

The licence, in one paragraph

MIT — commercial use, modification and redistribution with no revenue threshold, no territory clause and no naming requirement. The least encumbered licence of any model we stock.

What it needs

One configuration, the one we ship and quote: the Blackwell-optimised NVFP4 build, a Q8 cache at this model's native 1M window, plus the small runtime buffer.

Weights (NVFP4, 4-bit)	187.4 GB
Cache at 1M (Q8)	8.1 GB
Runtime buffer	+0.75 GB
Total	196.3 GB

Runs on: 4U4G-TURIN/HPR · 4UXGM-TURIN2 DIRECT · 8U16X-TURIN2 B300.

Not this one: NVIDIA RTX 5090 · NVIDIA DGX Spark (GB10) · NVIDIA RTX PRO 6000 Blackwell — smaller than this model at NVFP4 needs.

Not offered with it: NVIDIA DGX Station (GB300) — those machines are sold with the other model.

What we change, and why

We run the Blackwell-optimised NVFP4 build on NVIDIA hardware. NVFP4 is 4-bit in the tensor cores' native format on 5th-generation cards — the model keeps its quality and gains speed and memory, and it is the build our machines ship with.
Quantisation sized from measured model files, not nominal maths — the same rule as every model here.
KV cache at Q8_0 by default (half the cache, well under 1% quality cost), which is what keeps the 1M-token window affordable in memory.
We run it at its full native 1M window — nothing offered here is a scaled extension.
Prompts and tool configurations tuned per role, including Danish documents — the text-heavy roles that never need to see are exactly where this model earns its place.
No vision: it is paired with a seeing model for any job that has to read a document, and the blueprints say so rather than leaving it to be discovered.

Updates we have tracked

No tracked updates yet. Every update to this model is published in the news feed first and listed here in the same pass — one update, one news item, one line here.

Register

Nothing is on sale yet. Register for a role and you are told the day it opens — what people register for is what gets built first.