Project AURIONAI Use-Cases Research & Innovation
DeepSeek
LLM

DeepSeek-V4-Pro

DeepSeek's 1.6T-parameter MoE model released with MIT weights and code (no data). HF tree: finetunes 15, quants 27; 408k downloads/month.

DeepSeekDeepSeekChinaDeepSeek V4Family
Figure from the DeepSeek-V4-Pro model card.
Levelmodel
ClassLLM
Parameters1.6T
Context1,000,000
Released24 Apr 2026
Opennessopen weights permissive
Runs onhyperscale cluster, multi-GPU node
Statusactive

From the model's own page; who made each picture and where it comes from are listed on the Image Credits page.

Benchmark Results

vendor-reported
Codeforces ratingcoding
vendor-reported
3,206
02,0004,000

Score in rating. Higher is better.

blog.buildfastwithai.com, leaderboard.steel.dev
GPQA Diamondknowledge and reasoning
vendor-reported
90.1
050100

Score in %. Higher is better.

huggingface.co
Humanity's Last Exam (HLE)knowledge and reasoning
vendor-reported
37.7
050100

Score in %. Higher is better.

huggingface.co
LiveCodeBenchcoding
vendor-reported
93.5
050100

Score in %. Higher is better.

blog.buildfastwithai.com, leaderboard.steel.dev, huggingface.co
MMLU-Proknowledge and reasoning
vendor-reported
87.5
050100

Score in %. Higher is better.

huggingface.co
SWE-bench Verifiedcoding
vendor-reported
80.6
050100

Score in %. Higher is better.

blog.buildfastwithai.com, leaderboard.steel.dev, huggingface.co
Terminal-Bench 2.0agents and tool use
vendor-reported
67.9
050100

Score in %. Higher is better.

blog.buildfastwithai.com, leaderboard.steel.dev
Where this comes from
Every score names who ran it: vendor-reported, maintainer-run or independent. A rank is shown with the population it was ranked in (gap rule 3).

Lineage

successor
DeepSeek-V4-Pro to DeepSeek-V4-Pro-0813: successor, medium confidenceDeepSeek-V4-ProDeepSeek-V4-ProDeepSeek, Apr 2026DeepSeek-V4-Pro-0813DeepSeek-V4-Pro-0813DeepSeek, Aug 2026
DeepSeekDeepSeek-V4-Pro-0813 is a child (successor)medium confidencenotes
Where this comes from
Parent and child relations from model cards, Hugging Face base_model fields and the genealogy tables of the research notes. The tree shows two generations each way, 2 models. Confidence and evidence are the notes' own.

R&D Topics

Efficiency architectures (MoE, sparse attention, long context)produced byOpen-weight frontier releases as research substrateused by

Identity

Developer
DeepSeek
Family
DeepSeek V4
Released
24 Apr 2026 sourcessecondary report
Weights published
–
Record checked
10 Oct 2026
Where this comes from
Developer and dates from the research notes. Release date is the vendor's public announcement; weights date is when open weights were published.

Architecture and Size

Architecture
MoE Hybrid CSA + HCA attention, mHC, Muon optimiser; pretrained on over 32T tokens
Total parameters
1.6T sources
Active parameters
49B
Checkpoint parameters
–
Context length
1,000,000 tokens
Input
text
Output
text
Where this comes from
Parameter counts as the developer states them. Where the checkpoint holds a different total, both are shown (register A16).

Openness and Access

Openness
open weights permissive
Licence
MIT License sources
Access routes
open weights download
Where this comes from
The licence file governs over registry tags and aggregator labels.

Deployment

Runs on
hyperscale cluster, multi-GPU node
Footprint
T2 estimate (unverified). Output 384K. Preview release.
Runtimes
not in the notes
Where this comes from
Deployment tiers and runtimes as named in the notes. Footprint is the notes' figure, not a measurement by AURION.

Indonesia

Indonesian support
not stated
Regional languages
not in the notes
Jakarta hosting
unknown
Languages
not in the notes
Where this comes from
Indonesia fields are carried on every model; unknown is an allowed value.

Industry Use Cases

No industry use case links this model.

Sources

  1. DeepSeek V4-Pro Review: Benchmarks, Pricing & Architecture blog.buildfastwithai.com https://blog.buildfastwithai.com/deepseek-v4-pro-review-2026
  2. deepseek-ai (DeepSeek) huggingface.co https://huggingface.co/deepseek-ai
  3. leaderboard.steel.dev/leaderboards/swe-bench-verified.md https://leaderboard.steel.dev/leaderboards/swe-bench-verified.md
  4. DeepSeek Releases V4: Open-Source 1.6T MoE with 1M Context rits.shanghai.nyu.edu https://rits.shanghai.nyu.edu/ai/deepseek-releases-v4-open-source-1-6t-moe-with-1m-context
  5. deepseek-ai/DeepSeek-V4-Pro , Hugging Face huggingface.co https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "DeepSeek-V4-Pro", https://aurion.semai.id/models/deepseek-deepseek-v4-pro/, records of October 2026.