midnightcoderagent

/

MidnightCoder-80B

Like 1

Thank you very much for the feedback! I’m really glad you liked the model. 🙏

Regarding the metadata in the Midnight Coder CLI, one important note: it can mainly affect the context size before compression. As an alternative, you can manually define the context size using /context ctx_size.

To get started, I suggest these settings for coding:

• Temperature: 0.7
• Top K: 40
• Top P: 0.9
• Repeat Penalty: 1.05
• Presence Penalty: 0.0

These values are a good starting point, but they can be adjusted depending on the type of task. For more deterministic code generation, for example, a temperature between 0.2 and 0.5 may work better.

We’re also working on a new release of Midnight Coder, which will bring many improvements. The new version is becoming extremely powerful, especially in terms of coding-agent capabilities, context management, and efficiency with local models. 🚀

We’re also getting very good results even on older hardware, such as an 8 GB RX 580, running MoE models in the 35B parameter range. This version of the model will also be released soon.

Thank you again for testing it, for the feedback, and for the encouragement! We still have a lot ahead of us, but Midnight Coder is evolving rapidly. 💪🔥

Thank you very much for the feedback! I’m really glad you liked the model. 🙏

Regarding the metadata in the Midnight Coder CLI, one important note: it can mainly affect the context size before compression. As an alternative, you can manually define the context size using /context ctx_size.

To get started, I suggest these settings for coding:

• Temperature: 0.7
• Top K: 40
• Top P: 0.9
• Repeat Penalty: 1.05
• Presence Penalty: 0.0

These values are a good starting point, but they can be adjusted depending on the type of task. For more deterministic code generation, for example, a temperature between 0.2 and 0.5 may work better.

We’re also working on a new release of Midnight Coder, which will bring many improvements. The new version is becoming extremely powerful, especially in terms of coding-agent capabilities, context management, and efficiency with local models. 🚀

We’re also getting very good results even on older hardware, such as an 8 GB RX 580, running MoE models in the 35B parameter range. This version of the model will also be released soon.

Thank you again for testing it, for the feedback, and for the encouragement! We still have a lot ahead of us, but Midnight Coder is evolving rapidly. 💪🔥

thanks for the response. would there be any way to setup specific loading parameters if midnight coder is sending load instructions?
right now the models seem to load with default settings and default options. usually theres cfg files for that in which one can set per model parameters, including the context size u mentioned earlier.

Best Regards

Thank you for the suggestion!

Currently, all available configuration options can be accessed directly through the Midnight Coder TUI using the /command syntax. For example, you can use /context ctx_size to configure the context size.

For now, if you want to use specific default parameters for a model, you can configure them directly in LM Studio, llama.cpp, or Ollama, depending on the backend you are using.

Having model-specific configuration profiles for parameters such as temperature, top-k, top-p, and repeat penalty is definitely an interesting feature. We’ll evaluate the possibility of implementing this directly in Midnight Coder, with support across llama.cpp, Ollama, and LM Studio.

This would allow Midnight Coder to automatically apply optimized settings depending on the model and backend, instead of requiring users to configure them manually in each platform.

We’re already working on a new release of Midnight Coder with many improvements, so this is something we can evaluate for the upcoming versions.

Thanks again for the suggestion and for helping us improve Midnight Coder! 🚀💪

Best regards!