DeepSeek V4 Pro

Name: DeepSeek V4 Pro
Author: DeepSeek

Released

2026-04-24

Last refreshed

2026-10-08

Status

Researched 138d ago

Open sourceCommercial use: permittedCodingRAGAgentsLong contextClassificationJSON / Tool use

DeepSeek releases · 10 in the last 12 months · this family litChangelog →

Specifications

Family: DeepSeek V4
Released: 2026-04-24
Context: 1M
Max output: 384K
Parameters: 1.6T
Architecture: Mixture of Experts
Specialization: general
Openness: Open source
License: MITOSI-approvedPermitted
Weights: Available
Code: Unknown
Training: Pretrained

Created by

DeepSeek

Advancing artificial general intelligence (AGI).

Hangzhou, Zhejiang, China

Founded 2023

Website

Pricing

Output / 1M

$0.7266

Input / 1M

$0.3633

Cheapest of 6 routes · OpenRouter · cache read $0.0303

Providers(6)

DeepSeek Platform Fireworks AI OpenRouter Vercel AI Gateway Novita AI Featherless

View 6 provider routes

Links

Website HuggingFace

About

DeepSeek V4 Pro is DeepSeek's flagship open-weights model, released April 24 2026 under the MIT license. Architecture: 1.6T total / 49B active parameters, MoE with Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA) hybrid — requiring only 27% of inference FLOPs vs standard 1M-context transformers — plus Manifold-Constrained Hyper-Connections (mHC) and Muon Optimizer. Context window: 1,000,000 tokens; max output: 384,000 tokens (Think Max mode requires >=384K context). Text-only (no vision/image input). Supports three reasoning modes: Non-Think, Think High, Think Max. Function calling, tool use, and structured outputs supported.

Top use-case fit: coding, agents, and build tasks

Coding

Q/$ C

Agents

Q/$ B

Providers(6)

Compare all 6

Compare API pricing across 6 tracked provider routes for input and output tokens, batch, and cached reads when available.

Provider	Input / 1M	Output / 1M	Cache	Route
OpenRouter	$0.3633	$0.7266	read $0.0303	Serverless
Vercel AI Gateway	$0.435	$0.870	read $0.0036	Serverless
Featherless	$1.60	$3.20	-	Serverless
Novita AI	$1.60	$3.20	read $0.135	Serverless
Fireworks AI	$1.74	$3.48	-	Serverless
DeepSeek Platform	$1.32	$3.96	read $0.044	Serverless

Available via routers & gateways(4)

LiteLLM

Gateway

Open-source Python SDK and proxy server that unifies 100+ LLM APIs behind a single OpenAI-compatible interface, with load balancing, cost tracking, and configurable failover.

Free OSSDeepSeek Platform

OpenRouter

Hybrid

Unified hybrid gateway to 500+ models from 80+ providers via a single OpenAI-compatible API, with an optional Auto Router that selects a model per prompt from market spend data.

Passthrough + feeDeepSeek PlatformFireworks AI

Requesty

Hybrid

AI gateway and LLM router to 600+ models with routing and failover, caching, spend limits, governance, and observability.

Passthrough + feeDeepSeek Platform

Respan

Hybrid

Unified LLM engineering platform (gateway + observability + evals + prompt management) with one endpoint to 1,000+ models; previously Keywords AI.

SubscriptionDeepSeek Platform

Capabilities

ReasoningJSON / Tool useStructured OutputsPrompt Caching

Benchmark peer barsfor Coding

SWE-bench ProRank 40 of 62

93.0

89.9

81.3

81.2

DeepSeek V4 Procurrent

55.4

SWE-bench VerifiedRank 14 of 104

Claude Fable 5

96.0

Claude Opus 5

96.0

Claude Mythos Preview

93.9

Claude Opus 4.8

88.6

DeepSeek V4 Procurrent

80.6

HumanEvalRank 50 of 99

Claude Sonnet 4.6

98.0

96.7

Claude Opus 4.6

95.0

Grok-3

94.5

DeepSeek V4 Procurrent

76.8

Benchmark scores(17)

Scores are benchmark-specific and are direction-aware: the same numeric gap can mean very different outcomes across suites. Use the leaderboard context and this model's provider route to decide whether the winning margin is meaningful for your workload.

Benchmark	Score	Version	Evaluation	Source
Google-Proof Q&A	90.1	diamondObserved 2026-04-24	—	Source
Massive Multitask Language Understanding	90.1	5-shotObserved 2026-04-24	—	Source
MMLU PRO	87.5	Think Max mode (accuracy)Observed 2026-06-07	—	Source
SWE-bench Verified	80.6	SWE-bench VerifiedObserved 2026-04-24	—	Source
Chatbot Arena	1456.0	text (June 2026)Observed 2026-06-16	—	Source Source: LMArena (lmarena.ai), CC BY 4.0
LiveCodeBench	93.5	Think Max mode (pass@1)Observed 2026-06-07	—	Source
SWE-bench Pro	55.4	—Observed 2026-04-24	—	Source
HumanEval	76.8	Pass@1Observed 2026-04-24	—	Source
Mathematics Aptitude Test of Heuristics	64.5	—Observed 2026-04-24	—	Source
Terminal-Bench 2.0	59.1	independent evaluation (BenchLM, June 2026)Observed 2026-06-18	—	Source
SWE-bench Multilingual	76.2	—Observed 2026-04-24	—	Source
BrowseComp	83.4	BrowseComp (accuracy%)Observed 2026-06-07	—	Source
Humanity's Last Exam	37.7	HLE (accuracy)Observed 2026-06-07	—	Source
MCP-Atlas	69.4	independent evaluation (BenchLM, June 2026)Observed 2026-06-18	—	Source
GeneBench-Pro	2.4	xhighObserved 2026-06-30	—	Source
LMArena Text Arena - Creative Writing	1446.0	text creative_writing 2026-10-22Observed 2026-10-22	Configuration: DeepSeek-V4-Pro Harness: LMArena blind pairwise human votes, Bradley-Terry Evaluator: LMArena Confidence: confirmed Notes: LMArena Text arena, Creative Writing category (no style control), board published 2026-10-22 (HF lmarena-ai/leaderboard-dataset @a467e6b9, text config, latest split). Board entry 'DeepSeek-V4-Pro' (category creative_writing in text/latest-00000-of-00001.parquet): rank 37 of 416, rating 1446.2 (95% CI 1438.9-1453.4), 9731 votes. Source: LMArena (CC BY 4.0). Retrieved 2026-10-27.	Source Source: LMArena (lmarena.ai), CC BY 4.0
WMT26 General MT (human evaluation, cESA Avg. Rank)	17.5	WMT26 General MT findings, Figure 3 Avg. Rank (lower is better)Observed 2026-10-27	Configuration: DeepSeek V4 Pro Harness: Contrastive Error Span Annotation (cESA), human annotators Evaluator: WMT26 organizers Confidence: confirmed Notes: WMT26 General MT findings (2026.wmt-1.48), Figure 3: Avg. Rank 17.5 for the system listed as 'DeepSeek V4 Pro', averaged over the 23 language pairs it entered; rank 1 is best, so lower is better (not a 0-100 score). Per-pair cESA scores (0-100, higher is better, not comparable across pairs): cs-vi 77.0, cs-de 81.3, en-cs 56.6, en-de 66.7, en-is 71.4, en-zh-Hant 64.2, en-arz 80.0, en-ru 66.2, en-be 63.0, en-lld 22.9, en-se 43.8, en-et 71.8, en-ja 83.2, en-kk 74.3, en-id 78.9, zh-ja 72.8, en-ko 74.4, en-th 71.9, en-lij 20.1, cs-uk 82.2, en-uk 80.4, en-hy 67.3, en-zh-Hans 75.3. Source: WMT26 organizers. Retrieved 2026-10-27.	Source Source: Findings of the WMT26 General Machine Translation Shared Task (Kocmi et al., 2026)