DeepSeek-V3 (Mar 2025) TPS calculator
Each card below is assessed against this model at the context length and minimum quality you choose. Speed is an estimate for a single request, calculated from the card's memory bandwidth and the size of the model once compressed.
Calculated for this model
818 cards we hold specifications for
Which GPUs can run DeepSeek-V3 (Mar 2025)?
Set the inputs, read the answer
A longer conversation needs more memory, which can push this model off smaller cards.
Hides cards that would only fit the model by compressing it below this point.
0 cards match
Calculating| Needs | Quantisation | Fit | |||||
|---|---|---|---|---|---|---|---|
|
No card in our catalogue can run this model with these settings. |
|||||||
Speeds are estimates for a single request — one conversation at a time — calculated from memory bandwidth, model size and quantisation. Real throughput varies with the inference runtime and its version. Figures published by hardware vendors measure many simultaneous requests and are much higher.
On record
Full specification
Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.
Origin
Who built this model, where, and when it was published.
- Organisation
- DeepSeek
- Organisation type
- Industry
- Country
- China
- Published
- 24 March 2025
- Authors
- DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Haowei Zhang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Li, Hui Qu, J.L. C…
What it does
The problem areas the model was built for. A model can carry several of each.
- Domain
- Language
- Task
- Language modeling/generation, Code generation, Quantitative reasoning, Question answering
- Numerical format
- FP8
Size
How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.
- Parameters
- 671B
- Training data
- 14,800,000,000,000 tokens
- Batch size
- 62,914,560
Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token.
"We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities"
15360 * 4096
Training compute
The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.
- Training compute
- 3.3 × 10²⁴ FLOP
- How it was established
- Operation counting,Hardware
6 × 37B × 14.8T = 3.29e24 pretraining FLOPs, 1e22 posttrainining FLOPs - see doc below https://docs.google.com/document/d/1C4zBvHh4UNnzaorwQSdG61zQJsBV9X3i3A70NeelxHY/edit?usp=sharing
The training run
What it physically took to train: which chips, how many, for how long, and what that drew from the wall.
- Training hardware
- NVIDIA H800 SXM5
- Chips used
- 2,048
- Chip-hours
- 2,788,000
- Hardware utilisation
- MFU 19.5%
- Power draw
- 2.8 MW
- Compute cost
- $5,390,000
- Data centre
- Paper on DeepSeek-V3
Table 4: Training metric comparison "Causal MFU only takes into account the flops of the lower triangle of the attention matrix (in line with FlashAttention), while non-causal MFU includes the flops of the whole attention matrix (in line with Megatron)" TFLOPS (causal) = 385 H800 maximum FP8 performance = 1979 TFLOPS therefore MFU = 385/1979 = 0.1947
Availability
Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.
- Weights
- Open — downloadable
- Model access
- Open weights (restricted use)
- Training code
- Unreleased
- Hugging Face
- deepseek-ai
MIT and deepseek license https://github.com/deepseek-ai/DeepSeek-V3?tab=readme-ov-file I cannot see training code in this repo https://github.com/deepseek-ai/DeepSeek-V3?tab=readme-ov-file
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Likely above 10²³ FLOP
- Yes
- Why it is tracked
- Training cost
- Record confidence
- Confident
training cost was $5.3million USD (Table 1)
Sources
Where this record came from and when it was last checked.
- Reference
- DeepSeek-V3 Technical Report
- Last updated
- 19 February 2026
What the numbers mean
What you need to run it
DeepSeek-V3 (Mar 2025) reaches a parameter count of 671B. That is beyond what any single graphics card holds. Running it means either splitting it across several cards or renting hardware built for the job, and every card able to hold it alone is a datacentre part. The number that can: 0.
Background
DeepSeek-V3 (Mar 2025) was published by DeepSeek, in the country recorded as China, during March 2025. It comes out of an organisation categorised as industry.
It works in the domain of Language, and is recorded as performing the task of language modeling/generation, Code generation, Quantitative reasoning, Question answering.
The weights are published, so it can be downloaded and run on your own hardware indefinitely, offline, with no account attached. On Hugging Face it is published under the organisation deepseek-ai.
Training and provenance
The training run consumed about 3.3 × 10²⁴ FLOP, on hardware recorded as NVIDIA H800 SXM5. That figure measures what producing the model cost, and has no bearing on how fast it answers.
The training set ran to roughly 14,800,000,000,000 tokens of text.
It is tracked in the underlying dataset for one reason in particular: training cost.
Step by step
How to choose a GPU for DeepSeek-V3 (Mar 2025)
The table above has already assessed every card we hold specifications for against this model. Getting to your answer takes six steps.
-
01
Check what it needs before anything else
The table lists every card able to hold DeepSeek-V3 (Mar 2025). Capacity is the gate — a card either holds it or it does not.
-
02
Decide how long your conversations run
Set the context to what you will actually use. The cache grows with the conversation, and it is the usual reason a card that seemed fine stops fitting DeepSeek-V3 (Mar 2025).
-
03
Choose how far you will compress it
Each card runs the least-compressed copy it can hold. Setting a minimum quality drops the cards that only manage it by squeezing further than you would want, and holds the comparison at one level.
-
04
Compare tokens per second, not specifications
Sort by speed to see how cards rank for DeepSeek-V3 (Mar 2025). It will not match a gaming ordering, because generation is bound by memory bandwidth.
-
05
Check the fit verdict before buying
A tight fit runs, but leaves nothing spare for a longer conversation, in the case of DeepSeek-V3 (Mar 2025). Comfortable means you can grow the context later. That difference matters more than a few tokens per second, so buy for comfortable if you expect to.
-
06
Open the card you have settled on
Following a card through to its own page shows every other model it can hold, which is the question that follows once you have settled on DeepSeek-V3 (Mar 2025).
Answers
DeepSeek-V3 (Mar 2025) — common questions
DeepSeek-V3 (Mar 2025)— where can I download it?
Its weights are published on Hugging Face, under the organisation deepseek-ai. We do not host model files — this site calculates what hardware is needed to run them.
DeepSeek-V3 (Mar 2025)— how much compute was used to train it?
Training consumed around 3.3 × 10²⁴ FLOP, on hardware recorded as NVIDIA H800 SXM5. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.
DeepSeek-V3 (Mar 2025)— can I run it if it does not fit in my GPU?
Only by offloading, which is usually a false economy: the part held in system memory drags the whole thing down. The nearest miss we calculate falls short by 147.5 GB. Every figure here assumes the whole model is resident on the card.
DeepSeek-V3 (Mar 2025)— would two GPUs run it faster?
Capacity adds across cards; throughput does not. The number of cards already holding it on their own: 0. So a second card is rarely the answer here.
DeepSeek-V3 (Mar 2025)— why does the quantisation differ between cards?
Each card is shown running the least-compressed copy it can hold. The number of distinct compression levels across the cards that fit it: 1. Bigger cards get the more accurate version, and the quality floor above pins the comparison to one level.
DeepSeek-V3 (Mar 2025)— how accurate are these speed estimates?
Every figure is derived from memory bandwidth and model size, not benchmarked, which is why each is published as a range rather than a single number. One example: the range beneath each figure. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.
DeepSeek-V3 (Mar 2025)— is it open source?
Its weights are published, so it can be downloaded and run on your own hardware. Note that open weights is not the same as open source in the full sense — it says nothing about the training data, the training code, or the commercial terms attached.
DeepSeek-V3 (Mar 2025)— how many parameters does it have?
It has a parameter count of 671B. Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.
DeepSeek-V3 (Mar 2025)— who created it?
It was published by DeepSeek, based in China, an organisation categorised as industry.
DeepSeek-V3 (Mar 2025)— when was it released?
It was published in March 2025.
DeepSeek-V3 (Mar 2025)— what is it used for?
It works in the domain of Language, and is recorded as handling the task of language modeling/generation, Code generation, Quantitative reasoning, Question answering. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.
The other direction
Looking at it from the other side?
This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.