mradermacher

/

model_requests

Like 211

Hi @Rebis and @mradermacher ! I am the author of Mythos. Thank you so much for the interest, follow, and GGUF request!

A quick update on the model status:

• Architecture: Mythos is a custom Mixture-of-Experts (MoE) model with Liger Kernel Cross-Entropy designed for 500+ languages.
• Current Status: The model is actively undergoing pre-training from scratch (currently at Step 2,000+).
• GGUF Timeline: Because it is a custom MoE architecture and weights are still converging in pre-training, there are no final weights published yet. As soon as base training reaches convergence and we publish the Safetensors release, we will provide llama.cpp conversion scripts and GGUF quantizations!

If you'd like to track live training updates, telemetry, or chat about the architecture, you're very welcome to join our research Discord:
👉 https://discord.gg/KKVN5BShGj

Thank you for the support!

Hi,
I see, thanks for the information.
I'll look into signing up on Discord.
In any case, I am impressed by the speed of your responsiveness.

alright let me know then when it gets merged into main llamacpp and I will queue it =)

Thanks so much @RichardErkhov ! Really appreciate the support and looking forward to it.

Our base pre-training run and initial multilingual evaluations are estimated to take around 3 to 4 weeks to complete. Once the final weights are locked in and the llama.cpp conversion is ready, I'll definitely ping you here so you can queue the quantizations.

Thanks again! 🙌

Hi, for us to quant make sure that you are merged into main (ggeranov) llamacpp

Understood @RichardErkhov ! That makes total sense.

Once our base training finishes and we export the Hugging Face weights, we will ensure it aligns with llama.cpp main (or submit the architecture PR to ggerganov/llama.cpp) so your pipeline can convert it seamlessly.

Thanks for the heads-up and guidance! 🙌

or submit the architecture PR to ggerganov/llama.cpp

not or, have to. Otherwise we are not able to quant your model to avoid user confusion and the amount of llamacpp forks we have (we have 0 and want to maintain that number)

Understood and 100% agreed @RichardErkhov ! No forks at all.

When base training wraps, we will make sure the weights are formatted so they run directly on official ggerganov/llama.cpp main before pinging you.

Thanks for clarifying the policy! 👍