Models / DeepSeek/ DeepSeek V4 Pro

DeepSeek V4 Pro

Name: DeepSeek V4 Pro
Author: DeepSeek

DeepSeek · released Apr 22, 2026 · deepseek-ai/DeepSeek-V4-Pro

InputOutput

Type: Open weightsMIT License
Params: 1.6T
Context: 1M

37B active per word · about 786K words of context

Our take

Written Aug 3, 2026

DeepSeek V4 Pro is a large downloadable text model with a one-million-token request limit and a permissive MIT licence. It excels at coding and hard-prompt tasks on the independent leaderboards we track, though its creative-writing score lags behind its other skills.

Who should pick it

Choose this for coding-heavy workloads where its highest arena score matters, or for long-document analysis at one million tokens. It suits cost-sensitive inference through the cheapest hosts, and teams who need a permissive licence for commercial use or modification. Skip it if you need image, video or audio input, if creative writing is your main use, or if you want predictable speed across providers.

The case for it

Exceptional coding performance relative to its other capabilities: its coding score is 44 points above its overall text score on the arena leaderboard.
One-million-token request limit, rare among downloadable models, for long-document analysis.
Strong hard-prompt handling, scoring 23 points above its overall text score.
Permissive MIT licence allows commercial use, modification and redistribution.

The case against it

Creative writing is its weakest arena skill, scoring 56 points below its coding score and 12 points below its overall text score.
Throughput varies more than sixfold across providers, so speed depends heavily on which host you pick.
Text-only: no image, video or audio input or output.

How good is it?

What has been measured, on which board, and where DeepSeek V4 Pro placed.

IntelligencePuzzles, maths, exam questions

3.5 of 5

Arena Text (overall)28th of 143 · 1457.9

Arena Hard Prompts 28th of 143Arena Maths 36th of 139LiveBench Mathematics 13th of 35LiveBench Data Analysis 14th of 35LiveBench Reasoning 20th of 35

Also on this board: 1455.1 via Thinking (Jul 30, 2026). Read the pair, not the higher one.

CodingWriting and fixing code on its own

3.5 of 5

Arena Coding32nd of 143 · 1502.2

Arena Code (WebDev) 31st of 74LiveBench Coding 30th of 35

Also on this board: 1489.9 via Thinking (Jul 30, 2026). Read the pair, not the higher one.

AgenticPlanning, calling tools, staying on task

2.5 of 5

Arena Agent (IPS)21st of 36 · −0.008

LiveBench Agentic Coding 27th of 35

WritingWe do not rate this

Scored, not ratedThe placings are on the right.

Two boards come close and neither tests writing: Arena Creative Writing asks people which of two replies they prefer, and LiveBench Language tests whether a model understood a passage. So we show where DeepSeek V4 Pro placed and give it no mark out of five.

Arena Creative Writing 22nd of 143 · 1446.1LiveBench Language 16th of 35 · 78.1

Also scored, on boards we give no mark for

Arena Instruction Following 25th of 143LiveBench 24th of 35LiveBench Instruction Following 27th of 35

These tests check whether a model follows instructions — a precondition for all the work above, but not a measure of how well that work is done, which is why they get no rating.

Every published score for this model16 scoresEvery figure we hold, from 16 boards, with who ran it and a link to the source — including the boards no rating above is built on.

LiveBenchreasoning

71.6independentsource ↗

LiveBench Agentic Codingagentic

42.6independentsource ↗

LiveBench Codingcoding

70independentsource ↗

LiveBench Data Analysisreasoning

74.5independentsource ↗

LiveBench Instruction Followingchat

62.4independentsource ↗

LiveBench Languagechat

78.1independentsource ↗

LiveBench Mathematicsmath

90.7independentsource ↗

LiveBench Reasoningreasoning

82.7independentsource ↗

Arena Agent (IPS)agentic

−0.008independentsource ↗

Arena Codingcoding

1502.2independentsource ↗

Arena Creative Writingchat

1446.1independentsource ↗

Arena Hard Promptsreasoning

1480.9independentsource ↗

Arena Instruction Followingchat

1454.7independentsource ↗

Arena Mathsmath

1445.3independentsource ↗

Arena Text (overall)chat

1457.9independentsource ↗

Arena Code (WebDev)coding

1446independentsource ↗

Can you run it yourself?

Fits fully in memory on 0 of the 71 devices we track, at Q4.

Fits in memory: weights load entirely on the card
Spills to system RAM: some weights offload; much slower
Too large: will not load even with offload
est: size is calculated; the verdict could change by 10%

A card many people ownToo large

GeForce RTX 4090 · 24 GB

Weights at Q4_K_M1008 / 24 GBest

Usable contextNot calculated for spilled setups

Decode speedNot estimated for spilled setups

Too large for this card. The weights do not fit even with part of them offloaded to system memory.

One step upToo large

Apple M1 Pro (16-core GPU) · 32 GB

Weights at Q4_K_M1008 / 32 GBest

Usable contextNot calculated for spilled setups

Decode speedNot estimated for spilled setups

Too large for this card. The weights do not fit even with part of them offloaded to system memory.

One step downToo large

Radeon RX 7900 XT · 20 GB

Weights at Q4_K_M1008 / 20 GBest

Usable contextNot calculated for spilled setups

Decode speedNot estimated for spilled setups

Too large for this card. The weights do not fit even with part of them offloaded to system memory.

Your hardware

Checking your profile…

Memory use by level

Against a 24 GB card.

Q4_K_M

recommended

1008 GBest

Too large

Q5_K_M

1182.6 GBest

Too large

Q8_0

1766.7 GBest

Too large

All 3 sizes here are calculated, not measured. We hold no measured file for this model, so each size comes from the parameter count and every verdict above inherits that uncertainty.

Check against your own machine → · All 71 devices, with every size →

Or rent it from someone else

20 providers, from 23 listings.

Cheapest published offer

OpenRouter

Cheapest of 23 live listings. Picked at the widest standard context we hold, within one quantisation slice, so the numbers beside it are a price one host actually charges.

per 1M tokens: $0.43 in / $0.87 out
Context served: 1M
Throughput: Not measured

Current provider offers with price, context and prompt-privacy answers
Provider	In / out per 1M tokens	Context	Throughput	Trains on prompts	Logs prompts	Zero retention
OpenRouter	$0.43 / $0.87	1M	not measured	Unknown	Unknown	Unknown
DeepSeek	$0.43 / $0.87	1M	27 tok/s	Yes	Yesunknown period	Unknown
Baidufp8	$0.54 / $1.08	1M	61 tok/s	No	Yesunknown period	Unknown
StreamLakefp8	$0.61 / $1.22	1M	32 tok/s	No	Yesunknown period	Unknown
GMICloudfp8	$0.68 / $1.36	1M	29 tok/s	No	Yesunknown period	Unknown
Ionstreamfp4	$1.13 / $2.26	1M	14 tok/s	No	No	Confirmed
Novita AIfp8	$1.17 / $2.34	1M	38 tok/s	No	No	Confirmed
Waferfp4	$1.20 / $2.40	1M	not measured	No	No	Unknown
DeepInfrafp4	$1.30 / $2.60	1M	18 tok/s	No	No	Confirmed
DeepInfrafp4	$1.30 / $2.60	1M	not measured	Unknown	Unknown	Unknown
DigitalOcean Gradient	$1.39 / $2.78	262K	8 tok/s	No	No	Confirmed
Alibaba Cloudfp8	$1.42 / $2.83	1M	50 tok/s	No	Yesunknown period	Unknown
Alibaba Cloud	$1.42 / $2.83	1M	53 tok/s	No	Yesunknown period	Unknown
SiliconFlowfp8	$1.50 / $3.13	1M	40 tok/s	No	No	Confirmed
Novita AI	$1.60 / $3.20	1M	not measured	Unknown	Unknown	Unknown
Venice AI	$1.65 / $3.30	1M	43 tok/s	No	No	Confirmed
AtlasCloudfp4	$1.68 / $3.38	1M	36 tok/s	No	Yesunknown period	Unknown
Basetenfp4	$1.74 / $3.48	262K	113 tok/s	No	No	Confirmed
CoreWeavefp8	$1.74 / $3.48	1M	12 tok/s	No	No	Confirmed
Together AI	$1.74 / $3.48	512K	55 tok/s	No	No	Confirmed
Parasailfp8	$1.74 / $3.48	1M	47 tok/s	No	No	Confirmed
Fireworks AI	$1.74 / $3.48	1M	43 tok/s	No	No	Confirmed
Cloudflare Workers AI	$1.74 / $3.48	393K	36 tok/s	No	Yesunknown period	Unknown

Across the 23 listings we hold: 19 say they do not train on prompts, 1 say they do and 3 do not say. 11 appear in the zero-retention registry we check; the rest are unknown to us rather than confirmed either way.

What each host's API supports

From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.

✓
Supported: ✗
Not supported
Not published: host gave no parameter list

API features per host
Provider	Tool calling	JSON output	Strict schema
OpenRouter	✓	✓	✓
DeepSeek	✓	✓	✗
Baidufp8	✓	✓	✓
StreamLakefp8	✓	✓	✓
GMICloudfp8	✓	✓	✗
Ionstreamfp4	✓	✓	✓
Novita AIfp8	✓	✓	✗
Waferfp4
DeepInfrafp4	✓	✓	✓
DeepInfrafp4
DigitalOcean Gradient	✓	✓	✓
Alibaba Cloudfp8	✓	✓	✓
Alibaba Cloud	✓	✓	✓
SiliconFlowfp8	✓	✓	✗
Novita AI
Venice AI	✓	✓	✓
AtlasCloudfp4	✓	✓	✓
Basetenfp4	✓	✗	✗
CoreWeavefp8	✓	✓	✓
Together AI	✓	✓	✓
Parasailfp8	✓	✓	✓
Fireworks AI	✓	✓	✓
Cloudflare Workers AI	✓	✓	✓

Tool calling: 20 of 23 listings say yes, 3 publish no parameter list. JSON output: 19 of 23 listings say yes, 1 says no, 3 publish no parameter list. Strict schema: 15 of 23 listings say yes, 5 say no, 3 publish no parameter list.

Models people weigh against DeepSeek V4 Pro

Shortlist any of these to compare them side by side.

Weighed against itKimi K2.7 Code Kimi K2.6 Kimi K2.5

Compare them field by field →

When we formed this view

And whether anything has moved since.

Dates behind this page

Aug 3, 2026Price changeStreamLake raised DeepSeek V4 Pro pricing by 29%input +29% ($0.47 → $0.61 per 1M tokens); output +29% ($0.95 → $1.22 per 1M tokens); cache read +29% ($0.039 → $0.051 per 1M tokens)detected

Aug 2, 2026Price changeDeepSeek V4 Pro cut across 2 hosts, by up to 16% at BaiduDeepSeek V4 Pro moved on 2 hosts: Baidu: input −16% ($0.64 → $0.54 per 1M tokens); output −16% ($1.28 → $1.08 per 1M tokens); cache read −16% ($0.053 → $0.045 per 1M tokens); StreamLake: input −18% ($0.57 → $0.47 per 1M tokens); output −18% ($1.15 → $0.95 per 1M tokens); cache read −18% ($0.048 → $0.039 per 1M tokens)detected

Aug 2, 2026BenchmarkScored 1502.2 on Arena Codingleaderboardannounced

Aug 2, 2026BenchmarkScored 1446.1 on Arena Creative Writingleaderboardannounced

Aug 2, 2026BenchmarkScored 1480.9 on Arena Hard Promptsleaderboardannounced

Aug 2, 2026BenchmarkScored 1454.7 on Arena Instruction Followingleaderboardannounced

Aug 2, 2026BenchmarkScored 1445.3 on Arena Mathsleaderboardannounced

Aug 2, 2026BenchmarkScored 1457.9 on Arena Text (overall)leaderboardannounced

Aug 2, 2026BenchmarkScored 1446 on Arena Code (WebDev)leaderboardannounced

Aug 1, 2026Price changeDeepSeek V4 Pro repriced across 2 hosts, from a 3% rise at Baidu to a 7% cut at StreamLakeDeepSeek V4 Pro moved on 2 hosts: Baidu: input +3% ($0.63 → $0.64 per 1M tokens); output +3% ($1.25 → $1.28 per 1M tokens); cache read +3% ($0.052 → $0.053 per 1M tokens); StreamLake: input −7% ($0.67 → $0.62 per 1M tokens); output −7% ($1.34 → $1.25 per 1M tokens); cache read −7% ($0.056 → $0.052 per 1M tokens)detected

Prices last checked 5h ago

What we do not know about this model yet

We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
3 of 23 listings publish no parameter list, so what their API accepts is unknown to us.
Nothing we hold says whether an endpoint streams, so we do not show it either way.
3 of 23 listings do not say whether they train on prompts.

Licence and identifiers

What the licence allowsMIT License, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.

Licence

MIT License

permissiveCommercial use allowed

Fully permissive: do anything with attribution. No patent grant, unlike Apache-2.0.

Identifiers

Hugging Face: deepseek-ai/DeepSeek-V4-Pro
Architecture: Mixture of experts
Modality record: text->text
Catalogue slug: deepseek-deepseek-v4-pro

Machine-readable model card (omc.json) →

Explore furtherWhere to run DeepSeek V4 Pro Best open-weight models Longest context windows

Something wrong on this page? Tell us