DeepSeek dropped a new model this week, and my inbox has three variations of
the same question: "this is open weights, right, can I run it on my server?"
Short answer: technically yes, practically no. Let me walk through why.
What is DeepSeek V4.1 Flash
DeepSeek V4.1 Flash is a 552-billion-parameter, MIT-licensed mixture-of-
experts model released on 10 September 2026, which DeepSeek says beats GPT-5.6
Sol and Claude Opus-5.0 on several agentic coding benchmarks while activating
only 8-16 billion parameters per token. Weights are public on Hugging
Face, and the
official write-up is on DeepSeek's
site and the API docs
changelog.
Starting today, 14 September 2026, DeepSeek is also rerouting all
deepseek-v4-pro API traffic to V4.1-Flash, and phasing V4-Pro out entirely.
If you were calling the Pro endpoint, you are already on this new model
whether you asked for it or not.
The architecture, in plain words
This is not just a bigger version of the old V4-Flash. DeepSeek rebuilt the
internals. It uses a new "Causal Encoder-Decoder" (CED) design: a 40-layer
transformer split into a 20-layer causal encoder and a 20-layer decoder. There
are 384 routed experts plus 1 shared expert per MoE layer, and only 6 routed
experts fire per token.
It is natively multimodal, meaning it reads images directly without a bolted-
on vision encoder. Context window is 1 million tokens, with output up to 384K
tokens. Reasoning effort is a dial you can turn from 1 to 100, so you trade
speed for depth on demand.
The KV cache work is genuinely clever: about 890 bytes per token, roughly a
quarter of the old V4-Flash and an eighth of the SSD storage, using Compressed
Sparse Attention 2, FP4 (E2M1) KV caching, SWA Bounded Replay, and a
196-billion-parameter "Engram" memory module accessed sparsely. DeepSeek
claims a 437x cache reduction versus the original V1 model. It was trained
from scratch on 45 trillion tokens, with context extended to 1M around the
34T-token mark. If you've followed the Qwen3 architecture
previews, this is the same industry trend: shrink the active compute, keep
the total knowledge huge.
Where it actually wins (and where it doesn't)
The benchmark numbers are real, but they are mixed, not a clean sweep. Here is
the comparison that matters for coding and agent work, all at max reasoning
effort.
| Benchmark | DeepSeek V4.1 Flash | GPT-5.6 Sol | Claude Opus-5.0 |
|---|---|---|---|
| DeepSWE v1.1 | 74.2 | 73.0 | 74.0 |
| Terminal-Bench 2.1 | 90.6 | 88.8 | 89.1 |
| AutomationBench | 54.8 | 45.8 | 50.3 |
| Agent's Last Exam | 31.8 | 26.7 | 28.6 |
| Terminal-Bench 4.0 | 31.2 | n/a | 51.8 |
| ProgramBench | 20.3 | n/a | 37.0 |
| NL2Repo-Bench | 64.0 | n/a | 75.3 |
| ExploitGym | 15.3 | 33.7 | n/a |
So it genuinely leads on Terminal-Bench 2.1, AutomationBench, Agent's Last
Exam, CyberGym (88.1 vs Sol's 84.5), and coding contests (Codeforces rating
3471, GPQA Diamond 90.9). But on the newer, harder Terminal-Bench 4.0 and
ProgramBench, Opus-5.0 beats it by a wide margin, and on exploit-writing
(ExploitGym) GPT-5.6 Sol is well ahead. Independent testers on
flowtivity.ai
also found it failed a Rubik's Cube simulation stress test that other coding
models handle fine. This is not a model that beats everyone at everything. It
is strong on specific agentic tasks and weaker on longer, messier real-world
repo work.
The self-hosting reality
Here is the part that matters most for this blog. Yes, the weights are on
Hugging Face under MIT licence. No, that does not mean you can run this on
your homelab box.
The checkpoint is around 510 GB. That alone rules out almost every consumer
setup. But the bigger issue is something people misunderstand about sparse MoE
models: even though only 8-16B parameters are active per token, all 552B
parameters still need to sit in memory (VRAM, ideally), because the router can
send any token to any of the 384 experts. "Sparse" saves you compute, not
memory. You cannot page experts in and out from disk fast enough to keep up
with real inference speed.
For comparison, the previous-generation V4-Flash was 304B total, about 13B
active, a 166.9 GB download, and it fit on 2x H200 GPUs. This new model is
bigger, not smaller. Industry estimates from Yotta
Labs put the practical floor at a full 8-GPU node, not the 2x H200
that used to be enough. According to
mindstudio.ai,
DeepSeek has not published an official minimum-VRAM number, community GGUF and
AWQ quantizations were still pending at launch, and the early local runs
people reported were on multi-GPU servers or rented cloud GPUs, not gaming
rigs or single-box homelabs. There's also a good technical rundown of the KV
cache work on NYU Shanghai's
blog
if you want the deeper math.
In my lab, the practical cutoff for "can I run this on hardware I own" has
always been: does it fit on one or two consumer or prosumer GPUs after
quantization. This model doesn't get close. Even a heavily quantized version
would need well over 128 GB of fast memory just for weights, before you add KV
cache for a 1M-token context. That is enterprise GPU cluster territory, not a
Proxmox box in a spare room (my Proxmox
guide assumes normal
hardware, not 8-GPU nodes).
What I'd actually do
If you want to use V4.1-Flash today, use DeepSeek's API. Pricing is genuinely
cheap: roughly $0.15 per million input tokens off-peak (up to $0.30 at peak),
$0.60 per million output tokens, and as low as $0.003 per million for cache-
hit input. That is why they're killing V4-Pro, which cost about 4x more for
less capability.
For actual self-hosting on hardware a freelancer or small team can own, I'd
still point people toward smaller open models that fit on 1-2 GPUs, the kind I
covered in the Muse Glimmer writeup or the monthly open-source repos
roundup. If you're routing API calls to a model like this from your
own infrastructure, put it behind a private tunnel rather than exposing keys
on a public box; my WireGuard setup
guide
covers that, and if you're storing outputs or logs, something like Nextcloud
instead of Google Drive keeps that data under your control too.
My verdict
DeepSeek V4.1 Flash is a real technical achievement and the licence is
genuinely open. But "open weights" and "self-hostable" are two different
promises, and this release only keeps the first one for most of us. Unless you
have access to an 8-GPU enterprise node, you are using this model through
DeepSeek's API like everyone else, cheap pricing and all. That's not a bad
outcome, just don't let the word "open" fool you into thinking it belongs on
your home server.
Top comments (0)