Alireza Dadgarnia

[Alireza Dadgarnia @adotdad]

@SelfHostedSI @bnjmn_marie The paper’s number is for 8 H200s and my new number is for the same node.

[Alireza Dadgarnia @adotdad]

@SelfHostedSI @bnjmn_marie That’s something that belongs to the past! Now for a model like K2.5, even with much more complex formats like the IQ ones, GSQ takes about 5 hours.

[Michael Helcig @mhelcig]

Our largest model yet.

763B parameters, compressed with GSQ-RCO and co-optimized with Strata (by @coldniko) for local inference.

On Hugging Face this Friday.

Much more is coming. Follow us for future releases.

[Alireza Dadgarnia @adotdad]

This project is personally very important to me because I believe expert pruning is essential if we want to run much larger models on consumer setups. So we're continuing to push in this direction while also improving GSQ and RCO for our future quantized models.

[Alireza Dadgarnia @adotdad]

We've updated Qwen3.8-Flash-Next-GSQ-RCO-Coder! Improved RCO and more diverse calibration data boost its agentic coding and vision capabilities. We also pushed the model to its full 262K context window through multiple compactions without encountering any looping behavior.