Model index · DeepSeek · updated 2026-08-16

DeepSeek models and prices

10 models, 10 with a published price , 7 with open weights · released 2024-12-26 to 2026-08-12.

You can buy from DeepSeek directly, or from 28 other hosts

We hold offerings sold by DeepSeek itself as well as by independent hosts. Buying direct is not automatically the cheapest route, and it is not automatically the same product either — hosts differ on quantisation, served context window and rate limits. Publishes cache-hit and cache-miss input rates separately, and has used off-peak discounts.

What DeepSeek makes

CategoryModelsBrowse
Text 10 all Text models →

Cheapest DeepSeek models per 1M tokens

Standard on-demand tier, cheapest host per model. Media models are priced in their own unit and listed further down rather than mixed into this table.

ModelBlended /1MCheapest hostQualityContext
DeepSeek V4 Flash 0731 $0.093 DigitalOcean not measured 1.05M
DeepSeek V3.2 $0.241 StreamLake 60/100 164k
DeepSeek V3.2 Exp $0.305 Novita 60/100 164k
DeepSeek V3 0324 $0.405 DeepInfra not measured 164k
DeepSeek V3.1 $0.425 OpenRouter (default route) * not measured 164k
DeepSeek V3.1 Terminus $0.440 OpenRouter (default route) * 80/100 164k
DeepSeek V3 $0.450 OpenRouter (default route) * not measured 164k
DeepSeek V4 Pro $0.528 Baidu 100/100 1.05M
R1 Distill Llama 70B $0.800 OpenRouter (default route) * not measured 8k
R1 $1.15 OpenRouter (default route) * not measured 164k

Measured quality across DeepSeek models

Our own probe suite at temperature 0, scored by deterministic checks — no LLM judge and no vendor claim. Measured on request, so most of the catalog is unscored; a blank is not measured, never a low score. Confidence falls when probes could not be read, which happens on models that answer in a hidden reasoning channel.

ModelScoreTierRanking probesConfidence
DeepSeek V4 Pro 100/100 frontier 5/5 1
DeepSeek V3.1 Terminus 80/100 near-frontier 4/5 1
DeepSeek V3.2 60/100 mid 3/5 1
DeepSeek V3.2 Exp 60/100 mid 3/5 1

Open weights from DeepSeek

Parameter counts are computed from the published weight files, and licences are what the authors attached — both from the weights registry rather than from any seller. Licence class decides whether you can self-host commercially without signing something.

ModelParametersLicenceDownloads (30d)Repo
R1 684.53B mit 8.52M deepseek-ai/DeepSeek-R1
DeepSeek V4 Flash 0731 290.94B mit 2.28M deepseek-ai/DeepSeek-V4-Flash
DeepSeek V4 Pro 1598.84B mit 1.41M deepseek-ai/DeepSeek-V4-Pro
DeepSeek V3.2 685.4B mit 1.09M deepseek-ai/DeepSeek-V3.2
R1 Distill Llama 70B 70.55B mit 0.28M deepseek-ai/DeepSeek-R1-Distill-Llama-70B
DeepSeek V3.2 Exp 685.4B mit 0.18M deepseek-ai/DeepSeek-V3.2-Exp
DeepSeek V3.1 Terminus 684.53B mit 0.02M deepseek-ai/DeepSeek-V3.1-Terminus

Who sells DeepSeek models

Offering rows per seller across this vendor's models. A seller appearing many times serves many of the models, not the same one repeatedly.

SellerOfferings
Novita 17
OpenRouter (default route) 10
DeepInfra 7
SiliconFlow 7
AtlasCloud 6
StreamLake 5
Baidu 3
GMICloud 3
Venice 3
DigitalOcean 3
Alibaba 3
CoreWeave 3
Google 2
Phala 2
SambaNova 2
DeepSeek 2 the vendor itself
Cloudflare 2
Parasail 2
Fireworks 2
Friendli 1
Amazon (Bedrock) 1
Ionstream 1
BaseTen 1
Together 1
Ambient 1
Mancer 2 1
Mara 1
Azure 1
Crusoe 1

One API for every DeepSeek model

VernaOne fronts the sellers on this page behind a single interface, so switching model or host is a config change — with automatic fallback if quality or latency regresses.

Try VernaOne free →

All 10 DeepSeek models

Newest first, priced or not. A model with no price is a fact about what its sellers publish, not a gap we are hiding.

ModelCategoryReleasedPriceKindJob
DeepSeek V4 Pro Text 2026-08-12 $0.528 /1M reasoning
open weights
text generation
DeepSeek V4 Flash 0731 Text 2026-08-01 $0.093 /1M reasoning
open weights
text generation
DeepSeek V3.2 Text 2025-12-01 $0.241 /1M reasoning
open weights
text generation
DeepSeek V3.2 Exp Text 2025-09-29 $0.305 /1M reasoning
open weights
text generation
DeepSeek V3.1 Terminus Text 2025-09-22 $0.440 /1M reasoning
open weights
text generation
DeepSeek V3.1 Text 2025-08-21 $0.425 /1M reasoning text generation
R1 Text 2025-05-28 $1.15 /1M reasoning
open weights
text generation
DeepSeek V3 0324 Text 2025-03-24 $0.405 /1M standard text generation
R1 Distill Llama 70B Text 2025-01-23 $0.800 /1M reasoning
open weights
text generation
DeepSeek V3 Text 2024-12-26 $0.450 /1M standard text generation

Other companies

Alibaba · Black Forest Labs · LTX · Kling · OpenAI · Google · Stability AI · Minimax · Bria AI · Bytedance · Pixverse · Ideogram · Tencent · Luma AI · All models

Frequently asked

How many models does DeepSeek have?

We track 10 DeepSeek models across 1 category, 10 of them with a published price. The list is rebuilt from provider APIs, the vendor's own docs and the weights registry rather than maintained by hand, so new releases appear without us writing about them.

Can I buy DeepSeek models directly from DeepSeek?

Yes — we hold offerings sold by DeepSeek itself (as deepseek), alongside 28 other hosts. Buying direct is not automatically cheapest: each model page lists every seller with its price and the evidence behind it.

Are DeepSeek models open weights?

7 of 10 have a public weights repository. Licence mix: 7 open weights permissive. Licence class decides whether you can self-host commercially without a separate agreement, so check the model page before planning a deployment.