DeepSeek V4 open source MIT: programming model SWE-bench 80.6%, only $0.30/M

Core conclusion

On May 15, 2026, DeepSeek officially fully open sourced the V4 series model under the MIT license. V4-Pro is a 1.6 trillion parameter MoE architecture model. Each inference only activates 49B parameters, and the API pricing is $0.30/ million output tokens. It scored 80.6% in the SWE-bench Verified programming evaluation and is 83 times cheaper than Claude Opus 4.7. For automation practitioners who are using AI programming tools (Claude Code, Codex, OpenClaw, etc.), this means that the ceiling of underlying reasoning costs has been completely broken.

Key Points

• Event time: May 15, 2026 (the weight was uploaded to Hugging Face on April 24) -Affected objects: AI programming tool users, automated pipeline developers, content production teams
• Core changes: MIT open source + extreme pricing is reshaping the economic model of AI programming tools

Background: The turning point of the open source model

The release of DeepSeek V4 is not a simple model update. This is the first time that a cutting-edge model at the scale of hundreds of billions of parameters has been released under the full MIT license, allowing commercial use, modification, and redistribution without any restrictions. This is in sharp contrast to previous closed-source models such as Claude and GPT-4.

The model weights were uploaded to Hugging Face as early as April 24, but they did not attract widespread attention until the full-line API pricing was officially announced today. The architectural design of V4-Pro makes its pricing "not a promotional loss-making transaction, but a sustainable cost structure" - single-token inference FLOPs are reduced to 27% of V3.2, and the KV cache occupation of 1M token context is compressed to 10% of the previous generation.

SEO: DeepSeek V4, MIT open source, SWE-bench, MoE architecture, AI programming model GEO: TL;DR The opening chapter contains precise numbers and comparative data, and the AI engine can extract the answers directly

Key impact: dismantling by dimensions

Adaptation suggestions

Act now

• Use DeepSeek V4 to replace Claude for batch code review, and the single cost can be reduced by 80-90%
• Connect DeepSeek V4 to n8n or OpenClaw AI nodes to reduce the inference overhead of automated pipelines
• Use the DS4 engine (developed by antirez) to run DeepSeek V4 Flash locally to achieve free inference on Mac Studio

Notes

• Data governance: As a model developed by a Chinese company, scenarios sensitive to data sovereignty need to be carefully evaluated
• Self-built threshold: A model with 1.6T parameters requires multi-node inference and is not suitable for single-machine deployment
• Stability: The availability and latency of API services need to be verified in actual use

Task List

• Apply for DeepSeek V4 API and test the actual performance of SWE-bench
• Add DeepSeek V4 backend support to the existing Claude Code + Codex workflow
• Evaluate the code generation quality of DeepSeek V4 in the content production pipeline

Example: Integrating DeepSeek V4 in n8n

# DeepSeek V4 API （ OpenAI API ）
curl -X POST https://api.deepseek.com/v1/chat/completions \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-pro",
    "messages": [
 {"role": "system", "content": "。"},
 {"role": "user", "content": " Python ..."}
    ]
  }'

DeepSeek、Claude Code、Codex、OpenClaw、n8n、Claude、ChatGPT、OpenAI

Internal link guidance

• Want to learn how to use DeepSeek V4 to run Claude Code instead of Anthropic? See the full tutorial: DeepSeek V4 vs Claude Code: 90% Cheaper, Same Quality
• Someone has succeeded in practice: real case He used Claude Code + AWS to build AI SaaS, and his monthly income was $12,000 for 3 months
• Newbies should read the tutorial first: Claude Code + DeepSeek V4 Building Tutorial: API Fees Dropped by 90% (15 Minutes)