deepseek-ai/DeepSeek-V2
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepSeek-V2.git synced 2026-11-05 00:47:23 +00:00
Code Issues 90 Projects Packages Wiki Activity

[GH-ISSUE #111] Significant Performance Regression in Code Editing: Request to Restore DeepSeek-Coder-V2 as Primary Coder Model #100

New issue
Open
opened 2026-10-26 21:07:53 +00:00 by bsakaguchi · 0 comments
bsakaguchi commented 2026-10-26 21:07:53 +00:00
Owner
Copy link

Originally created by @Victorcorcos on GitHub (Jun 20, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V2/issues/111

<html><head></head>

Significant Performance Regression in Code Editing: Request to Restore DeepSeek-Coder-V2 as Primary Coder Model

Summary

The merge of DeepSeek-Coder-V2 and DeepSeek-V2-Chat into DeepSeek-V2.5 has resulted in a dramatic performance regression for code editing tasks. Can I request that the deepseek/deepseek-coder API endpoint be redirected back to the original DeepSeek-Coder-V2 model rather than the merged V2.5 version?

Performance Impact

Aider Benchmark Results

The regression is most evident in the Aider code editing benchmark, which evaluates LLMs' ability to modify existing code:

  • DeepSeek-Coder-V2 (original): 73.7% pass rate - 1st position on leaderboard
  • DeepSeek-V2.5 (current deepseek/deepseek-coder): 17.8% pass rate - Significant drop in ranking

This represents a 76% decrease in performance on code editing tasks.

Benchmark Comparison

Model Aider Score Ranking Performance Change
DeepSeek-Coder-V2 (original) 73.7% 1st Baseline
DeepSeek-V2.5 (merged) 17.8% Low -76% ↓

Current Issue

When developers use aider --model deepseek/deepseek-coder, they expect to get the best coding-focused model from DeepSeek. However, they're now receiving the merged V2.5 model, which has significantly degraded code editing capabilities compared to the original specialized Coder-V2 model.

Impact on Developer Experience

  1. Unexpected Performance Drop: Developers who were using DeepSeek Coder V2 for its superior code editing capabilities are now experiencing much worse results without any warning
  2. Tool Integration Issues: Code editing tools like Aider that recommended DeepSeek Coder V2 as a top choice now perform poorly with the merged model
  3. Loss of Specialized Capabilities: The original Coder V2's specialized code editing abilities appear to have been diluted in the general-purpose merged model

Proposed Solutions

Option 1 (Preferred): Restore Original Coder Model

  • Redirect deepseek/deepseek-coder API endpoint back to DeepSeek-Coder-V2-Instruct
  • Keep DeepSeek-V2.5 available as deepseek/deepseek-chat for general-purpose tasks
  • This maintains the principle that specialized models should excel in their domains

Option 2: Provide Clear Model Differentiation

  • Create a new endpoint like deepseek/deepseek-coder-v2-original for the original model
  • Update documentation to clearly explain the performance differences
  • Provide migration guidance for users who need the original performance

Option 3: Improve V2.5 Code Editing Performance

  • Address the specific regression in code editing capabilities in the V2.5 model
  • Ensure the merged model doesn't sacrifice specialized performance for generality

Technical Context

According to your official documentation, DeepSeek-V2.5 is described as a "powerful combination" that "retains the robust code processing power of the Coder model." However, empirical testing shows this is not the case for code editing specifically, where the original Coder V2 significantly outperformed the merged version.

The Aider benchmark specifically tests:

  • Code modification and editing capabilities
  • Ability to understand and apply code changes accurately
  • Performance on real-world coding assistance scenarios

Request

I respectfully request that DeepSeek consider restoring the original DeepSeek-Coder-V2 as the model served by the deepseek/deepseek-coder endpoint, or at minimum, provide a clear path for developers to access the original high-performance code editing model.

This would:

  • Restore trust in DeepSeek's commitment to specialized model performance
  • Maintain compatibility for existing tools and workflows
  • Ensure developers get the best coding assistance experience

References

Thank you for considering this request. DeepSeek-Coder-V2 was an exceptional model for code editing, and I hope to see its capabilities restored for developers who depend on high-quality code assistance.

</html># Significant Performance Regression in Code Editing: Request to Restore DeepSeek-Coder-V2 as Primary Coder Model

Summary

The merge of DeepSeek-Coder-V2 and DeepSeek-V2-Chat into DeepSeek-V2.5 has resulted in a dramatic performance regression for code editing tasks. I request that the deepseek/deepseek-coder API endpoint be redirected back to the original DeepSeek-Coder-V2 model rather than the merged V2.5 version.

Performance Impact

Aider Benchmark Results

The regression is most evident in the Aider code editing benchmark, which evaluates LLMs' ability to modify existing code:

  • DeepSeek-Coder-V2 (original): 73.7% pass rate - #1 position on leaderboard
  • DeepSeek-V2.5 (current deepseek/deepseek-coder): 17.8% pass rate - Significant drop in ranking

This represents a 76% decrease in performance on code editing tasks.

Benchmark Comparison

Model Aider Score Ranking Performance Change
DeepSeek-Coder-V2 (original) 73.7% #1 Baseline
DeepSeek-V2.5 (merged) 17.8% Low -76% ↓

Current Issue

When developers use aider --model deepseek/deepseek-coder, they expect to get the best coding-focused model from DeepSeek. However, they're now receiving the merged V2.5 model, which has significantly degraded code editing capabilities compared to the original specialized Coder-V2 model.

Impact on Developer Experience

  1. Unexpected Performance Drop: Developers who were using DeepSeek Coder V2 for its superior code editing capabilities are now experiencing much worse results without any warning
  2. Tool Integration Issues: Code editing tools like Aider that recommended DeepSeek Coder V2 as a top choice now perform poorly with the merged model
  3. Loss of Specialized Capabilities: The original Coder V2's specialized code editing abilities appear to have been diluted in the general-purpose merged model

Proposed Solutions

Option 1 (Preferred): Restore Original Coder Model

  • Redirect deepseek/deepseek-coder API endpoint back to DeepSeek-Coder-V2-Instruct
  • Keep DeepSeek-V2.5 available as deepseek/deepseek-chat for general-purpose tasks
  • This maintains the principle that specialized models should excel in their domains

Option 2: Provide Clear Model Differentiation

  • Create a new endpoint like deepseek/deepseek-coder-v2-original for the original model
  • Update documentation to clearly explain the performance differences
  • Provide migration guidance for users who need the original performance

Option 3: Improve V2.5 Code Editing Performance

  • Address the specific regression in code editing capabilities in the V2.5 model
  • Ensure the merged model doesn't sacrifice specialized performance for generality

Technical Context

According to the Deep Seek official documentation, DeepSeek-V2.5 is described as a "powerful combination" that "retains the robust code processing power of the Coder model." However, empirical testing shows this is not the case for code editing specifically, where the original Coder V2 significantly outperformed the merged version.

The Aider benchmark specifically tests:

  • Code modification and editing capabilities
  • Ability to understand and apply code changes accurately
  • Performance on real-world coding assistance scenarios

Request

I respectfully request that DeepSeek consider restoring the original DeepSeek-Coder-V2 as the model served by the deepseek/deepseek-coder endpoint, or at minimum, provide a clear path for developers to access the original high-performance code editing model.

This would:

  • Restore trust in DeepSeek's commitment to specialized model performance
  • Maintain compatibility for existing tools and workflows
  • Ensure developers get the best coding assistance experience

References

Thank you for considering this request. DeepSeek-Coder-V2 was an exceptional model for code editing, and I hope to see its capabilities restored for developers who depend on high-quality code assistance.

Originally created by @Victorcorcos on GitHub (Jun 20, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V2/issues/111 <html><head></head><body><h1>Significant Performance Regression in Code Editing: Request to Restore DeepSeek-Coder-V2 as Primary Coder Model</h1> <h2>Summary</h2> <p>The merge of DeepSeek-Coder-V2 and DeepSeek-V2-Chat into DeepSeek-V2.5 has resulted in a dramatic performance regression for code editing tasks. Can I request that the <code>deepseek/deepseek-coder</code> API endpoint be redirected back to the original DeepSeek-Coder-V2 model rather than the merged V2.5 version?</p> <h2>Performance Impact</h2> <h3>Aider Benchmark Results</h3> <p>The regression is most evident in the Aider code editing benchmark, which evaluates LLMs' ability to modify existing code:</p> <ul> <li><strong>DeepSeek-Coder-V2 (original)</strong>: 73.7% pass rate - <strong>1st position</strong> on leaderboard</li> <li><strong>DeepSeek-V2.5 (current deepseek/deepseek-coder)</strong>: 17.8% pass rate - <strong>Significant drop in ranking</strong></li> </ul> <p>This represents a <strong>76% decrease in performance</strong> on code editing tasks.</p> <h3>Benchmark Comparison</h3> Model | Aider Score | Ranking | Performance Change -- | -- | -- | -- DeepSeek-Coder-V2 (original) | 73.7% | 1st | Baseline DeepSeek-V2.5 (merged) | 17.8% | Low | -76% ↓ <h2>Current Issue</h2> <p>When developers use <code>aider --model deepseek/deepseek-coder</code>, they expect to get the best coding-focused model from DeepSeek. However, they're now receiving the merged V2.5 model, which has significantly degraded code editing capabilities compared to the original specialized Coder-V2 model.</p> <h2>Impact on Developer Experience</h2> <ol> <li><strong>Unexpected Performance Drop</strong>: Developers who were using DeepSeek Coder V2 for its superior code editing capabilities are now experiencing much worse results without any warning</li> <li><strong>Tool Integration Issues</strong>: Code editing tools like Aider that recommended DeepSeek Coder V2 as a top choice now perform poorly with the merged model</li> <li><strong>Loss of Specialized Capabilities</strong>: The original Coder V2's specialized code editing abilities appear to have been diluted in the general-purpose merged model</li> </ol> <h2>Proposed Solutions</h2> <h3>Option 1 (Preferred): Restore Original Coder Model</h3> <ul> <li>Redirect <code>deepseek/deepseek-coder</code> API endpoint back to DeepSeek-Coder-V2-Instruct</li> <li>Keep DeepSeek-V2.5 available as <code>deepseek/deepseek-chat</code> for general-purpose tasks</li> <li>This maintains the principle that specialized models should excel in their domains</li> </ul> <h3>Option 2: Provide Clear Model Differentiation</h3> <ul> <li>Create a new endpoint like <code>deepseek/deepseek-coder-v2-original</code> for the original model</li> <li>Update documentation to clearly explain the performance differences</li> <li>Provide migration guidance for users who need the original performance</li> </ul> <h3>Option 3: Improve V2.5 Code Editing Performance</h3> <ul> <li>Address the specific regression in code editing capabilities in the V2.5 model</li> <li>Ensure the merged model doesn't sacrifice specialized performance for generality</li> </ul> <h2>Technical Context</h2> <p>According to your official documentation, DeepSeek-V2.5 is described as a "powerful combination" that "retains the robust code processing power of the Coder model." However, empirical testing shows this is not the case for code editing specifically, where the original Coder V2 significantly outperformed the merged version.</p> <p>The Aider benchmark specifically tests:</p> <ul> <li>Code modification and editing capabilities</li> <li>Ability to understand and apply code changes accurately</li> <li>Performance on real-world coding assistance scenarios</li> </ul> <h2>Request</h2> <p>I respectfully request that DeepSeek consider restoring the original DeepSeek-Coder-V2 as the model served by the <code>deepseek/deepseek-coder</code> endpoint, or at minimum, provide a clear path for developers to access the original high-performance code editing model.</p> <p>This would:</p> <ul> <li>Restore trust in DeepSeek's commitment to specialized model performance</li> <li>Maintain compatibility for existing tools and workflows</li> <li>Ensure developers get the best coding assistance experience</li> </ul> <h2>References</h2> <ul> <li><a href="https://aider.chat/docs/leaderboards/">Aider LLM Leaderboard</a> - Current benchmark results</li> <li><a href="https://api-docs.deepseek.com/news/news0905">DeepSeek-V2.5 Documentation</a> - Official merge announcement</li> <li>Original DeepSeek-Coder-V2 research showing 73.7% Aider performance</li> </ul> <hr> <p>Thank you for considering this request. DeepSeek-Coder-V2 was an exceptional model for code editing, and I hope to see its capabilities restored for developers who depend on high-quality code assistance.</p></body></html># Significant Performance Regression in Code Editing: Request to Restore DeepSeek-Coder-V2 as Primary Coder Model ## Summary The merge of DeepSeek-Coder-V2 and DeepSeek-V2-Chat into DeepSeek-V2.5 has resulted in a dramatic performance regression for code editing tasks. I request that the `deepseek/deepseek-coder` API endpoint be redirected back to the original DeepSeek-Coder-V2 model rather than the merged V2.5 version. ## Performance Impact ### Aider Benchmark Results The regression is most evident in the Aider code editing benchmark, which evaluates LLMs' ability to modify existing code: - **DeepSeek-Coder-V2 (original)**: 73.7% pass rate - **#1 position** on leaderboard - **DeepSeek-V2.5 (current deepseek/deepseek-coder)**: 17.8% pass rate - **Significant drop in ranking** This represents a **76% decrease in performance** on code editing tasks. ### Benchmark Comparison | Model | Aider Score | Ranking | Performance Change | |-------|-------------|---------|-------------------| | DeepSeek-Coder-V2 (original) | 73.7% | #1 | Baseline | | DeepSeek-V2.5 (merged) | 17.8% | Low | -76% ↓ | ## Current Issue When developers use `aider --model deepseek/deepseek-coder`, they expect to get the best coding-focused model from DeepSeek. However, they're now receiving the merged V2.5 model, which has significantly degraded code editing capabilities compared to the original specialized Coder-V2 model. ## Impact on Developer Experience 1. **Unexpected Performance Drop**: Developers who were using DeepSeek Coder V2 for its superior code editing capabilities are now experiencing much worse results without any warning 2. **Tool Integration Issues**: Code editing tools like Aider that recommended DeepSeek Coder V2 as a top choice now perform poorly with the merged model 3. **Loss of Specialized Capabilities**: The original Coder V2's specialized code editing abilities appear to have been diluted in the general-purpose merged model ## Proposed Solutions ### Option 1 (Preferred): Restore Original Coder Model - Redirect `deepseek/deepseek-coder` API endpoint back to DeepSeek-Coder-V2-Instruct - Keep DeepSeek-V2.5 available as `deepseek/deepseek-chat` for general-purpose tasks - This maintains the principle that specialized models should excel in their domains ### Option 2: Provide Clear Model Differentiation - Create a new endpoint like `deepseek/deepseek-coder-v2-original` for the original model - Update documentation to clearly explain the performance differences - Provide migration guidance for users who need the original performance ### Option 3: Improve V2.5 Code Editing Performance - Address the specific regression in code editing capabilities in the V2.5 model - Ensure the merged model doesn't sacrifice specialized performance for generality ## Technical Context According to the Deep Seek official documentation, DeepSeek-V2.5 is described as a "powerful combination" that "retains the robust code processing power of the Coder model." However, empirical testing shows this is not the case for code editing specifically, where the original Coder V2 significantly outperformed the merged version. The Aider benchmark specifically tests: - Code modification and editing capabilities - Ability to understand and apply code changes accurately - Performance on real-world coding assistance scenarios ## Request I respectfully request that DeepSeek consider restoring the original DeepSeek-Coder-V2 as the model served by the `deepseek/deepseek-coder` endpoint, or at minimum, provide a clear path for developers to access the original high-performance code editing model. This would: - Restore trust in DeepSeek's commitment to specialized model performance - Maintain compatibility for existing tools and workflows - Ensure developers get the best coding assistance experience ## References - [[Aider LLM Leaderboard](https://aider.chat/docs/leaderboards/)](https://aider.chat/docs/leaderboards/) - Current benchmark results - [[DeepSeek-V2.5 Documentation](https://api-docs.deepseek.com/news/news0905)](https://api-docs.deepseek.com/news/news0905) - Official merge announcement - Original DeepSeek-Coder-V2 research showing 73.7% Aider performance --- Thank you for considering this request. DeepSeek-Coder-V2 was an exceptional model for code editing, and I hope to see its capabilities restored for developers who depend on high-quality code assistance.
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
No results found.
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepSeek-V2#100
No description provided.