DeepSeek V4 Pro 0813

Flagship-scale agentic model with three reasoning effort levels and speculative decoding

Try now

read docs

About model

DeepSeek V4 Pro 0813 is the official release of DeepSeek's flagship model, superseding the V4 Pro preview with greatly enhanced agentic capabilities and performance gains that are especially pronounced in production environments. It is built on the V4 Pro preview's Mixture-of-Experts structure, 1.6T total parameters with 49B active per token, with a DSpark speculative decoding module attached for faster generation. Reasoning effort is adjustable per request across low, high, and max levels, letting the same deployment serve quick responses and deep deliberation. Across DeepSeek's published benchmark suite the release improves on the preview everywhere it is measured, with the largest jumps on repository-scale engineering and multi-tool agent tasks. Released under the MIT license. Available on Together AI.

‍

Total Parameters (49B activated)

1.6T

DeepSeek's flagship MoE architecture

Reasoning Effort Levels

Low, high, and max deliberation, set per request

Context Window

1.05M

Million-token context with hybrid attention built for efficient long-context inference

Model key capabilities

Production Agentic Work: The official release concentrates its largest gains on agentic tasks in production environments, superseding the preview
Adjustable Reasoning Effort: Low, high, and max effort levels set per request, trading response speed against reasoning depth on one deployment
Efficient Inference at Scale: DSpark speculative decoding module attached to the released checkpoint for faster generation
Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure

Performance benchmarks

Model	FrontierMath Tier 4	GPQA Diamond	HLE	SciCode	GDPval-AA	Terminal-Bench 2.1	Agent Arena	FrontierCode	DeepSWE
DeepSeek V4 Pro 0813	2.4%		42.7	50%	40%	64%	-2.9pp	17.6%
Related open-source models
Competitor closed-source models
Claude Fable 5	87.8%	92.6%	53%	60%	62%	85%		53.5%	70%
Claude Opus 5	73.2%	93.2%	53%	56%	68%	89%	+16.4pp	53.4%	74%
GPT-5.6 Sol	82.9%	94.1%	47%	56%	61%	88%		47.5%	73%
Grok 4.5	24.4%	93.1%	40%	54%	51%	82%	+4.2pp	42.4%	54%
GPT-5.6 Luna	61.0%	91.1%	37%	53%	54%	81%		39.8%	67%

API usage

cURL
Python
Typescript

Endpoint:

deepseek-ai/DeepSeek-V4-Pro-0813

curl -X POST "https://api.together.xyz/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-ai/DeepSeek-V4-Pro-0813",
    "messages": [
      {
        "role": "user",
        "content": "What are some fun things to do in New York?"
      }
    ]
}'

from together import Together

client = Together()

response = client.chat.completions.create(
  model="deepseek-ai/DeepSeek-V4-Pro-0813",
  messages=[
    {
      "role": "user",
      "content": "What are some fun things to do in New York?"
    }
  ]
)
print(response.choices[0].message.content)

import Together from 'together-ai';
const together = new Together();

const completion = await together.chat.completions.create({
  model: 'deepseek-ai/DeepSeek-V4-Pro-0813',
  messages: [
    {
      role: 'user',
      content: 'What are some fun things to do in New York?'
     }
  ],
});

console.log(completion.choices[0].message.content);

Model card
Architecture Overview:
• Mixture-of-Experts model with 1.6T total parameters and 49B activated per token, built on the DeepSeek V4 Pro preview structure with DeepSeekMoE feed-forward layers and Multi-Token Prediction retained
• Hybrid attention combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), requiring only 27% of single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2 at the longest context settings
• Manifold-Constrained Hyper-Connections (mHC) strengthen residual signal propagation across layers
• DSpark speculative decoding module attached to the released checkpoint
• Reasoning effort adjustable across low, high, and max levels

Training Methodology:
• Official dated release of the V4 Pro line, superseding the preview checkpoint
• Trained with the Muon optimizer for the majority of parameters, with AdamW retained for embeddings, the prediction head, and normalization weights
• DeepSeek reports the enhancements are especially pronounced in production environments
• Published agentic results evaluated at max reasoning effort with temperature 1.0 and top-p 0.95

Performance Characteristics:
• Expert knowledge (Humanity's Last Exam): 42.7, or 60.0 with tools, up from 37.7 and 48.2 for the preview
• Terminal coding (Terminal Bench 2.1, completing real tasks in a command-line environment): 87.9, up from 72.1
• Repository generation (NL2Repo, building full repositories from natural language specs): 61.5, up from 38.5
• Repository-scale software engineering (DeepSWE): 62.7, up from 12.8 for the preview
• Security engineering (Cybergym): 83.3, up from 52.7
• Multi-tool agent tasks (Toolathlon-Verified): 74.1, up from 55.9; Agents' Last Exam: 25.7; AutomationBench Public: 31.8, up from 12.8
‍
Prompting
Together AI API Access:
• Access DeepSeek V4 Pro 0813 via Together AI APIs using the endpoint deepseek-ai/DeepSeek-V4-Pro-0813
• Authenticate using your Together AI API key in request headers
• Runs in thinking mode by default, with a non-thinking mode available; set reasoning_effort to low, high, or max to control deliberation depth
• Recommended sampling: temperature 1.0, with top-p 0.95 for agentic workloads and 1.0 otherwise
• At high and max effort, allow generous output budgets: DeepSeek recommends up to 384K output tokens
• Available on Together AI serverless and dedicated infrastructure
‍
Applications & use cases
Production Agent Systems:
• Run multi-tool agent workflows where the release's largest gains are concentrated
• Tune reasoning effort per step: low for routine calls, max for hard planning
• Sustain long agent trajectories with generous output budgets at high effort

Repository-Scale Engineering:
• Generate and modify entire repositories from natural language specifications
• Run terminal-based agent sessions that complete real environment tasks end to end
• Apply the model to security-engineering and vulnerability-analysis workflows

Long-Context Analysis:
• Hold entire codebases and large document sets in the million-token context window
• Cross-reference material across a session without external retrieval for mid-sized corpora
• Keep long-running agent state in context across extended multi-turn sessions
‍

Related models

Model specifications

Model data

Model provider
DeepSeek
Type
Reasoning
Chat
Code
LLM
Main use cases
Reasoning
Features
Function Calling
JSON Mode
Intelligence
High
Deployment
Serverless
Dedicated
Endpoint
deepseek-ai/DeepSeek-V4-Pro-0813
Parameters
1.6T
Activated parameters
49B
Context length
1.05M
Input price
$1.32 / 1M tokens
$0.13 (cached)/1M
Output price
$3.96 / 1M tokens
Input modalities
Text
Output modalities
Text

Released
August 12, 2026
External link
Provider docs
Category
Chat

Run in Playground

Quickstart docs

Deploy model

DeepSeek V4 Pro 0813

About model

API usage

Model card

Prompting

Applications & use cases