looking forward for the release!!
super excited to see this new large MoE, looking forward for the release and the tech report!! :)
So am I. A.X 3 was such a unique and amazing model. A.X 4 despite being based on Qwen 2.5 improved on it by so much it felt unique and I loved using it. I'm so excited for the A.X-K1 release on 4th of January 2026. This is their first huge model and they already proved what they are capable of so I can only imagine how great it will be. The model card already looks super promising.
Thank you all for the interest and thoughtful feedback on the model — we truly appreciate it.
We wanted to share a quick update that the release schedule has been moved up, and the model is currently being uploaded today. The upload is expected to complete within the next couple of hours.
Thank you again for your interest and patience.
Thank you so much for releasing this amazing model!
With this release you provided the AI community a model of incredible value. It's rare to see foundational models of this size getting released. It being trained from ground up will probably make it as unique as A.X-3.1. Your training data seems to differ quite a lot from what other companies use which made A.X-3.1 one of my favorite models from the multiple hundred models I've tested so far. I've read under https://news.sktelecom.com/en/2533 that it can be used as a teacher model to act as a provider of knowledge. As far I'm aware this is the first massive teacher model that was ever publicly released. This makes it extremely valuable as this allows knowledge transfer to smaller models. I generally really appreciate the release of a knowledge focused model as knowledge was one of the areas, I felt many recently released models got worse maybe due to overfitting. With 519B this model has the perfect size as competitors such as DeepSeek (685B), Mistral-Large-3 (675B), Kimi-K2 (1T) are simply too large to efficiently run using affordable hardware at decent quality. Once someone implements llama.cpp support for it running this model on consumer hardware using 512 GiB of RAM at Q5_K_M, 256 GiB of RAM at IQ3_XXS or two PCs with 128 GiB of RAM each using RPC with decent context size and at reasonable speed thanks to the sparse MoE architecture used will become a reality.
I tried running it locally using the latest development version of vLLM but unfortunately, I was unable to locate any Transformers version with support for the AXK1ForCausalLM architecture or any pull request that would add support for it. A plain text search over the entirety of GitHub did not result in a single repository containingAXK1ForCausalLM. No rush I will still be as enthusiastic as I am now to try this model once there is a way to run it locally. I appreciate that the release schedule has been moved up as this allows me to prepare the infrastructure to run this efficiently once local inference support for it gets released.
Thank you for the detailed feedback — we really appreciate your interest and thoughtful analysis.
Regarding vLLM support: the model is already working well in our internal tests. Public vLLM support is currently being finalized, and we plan to submit a pull request in the coming days.
Thanks again for your enthusiasm and patience.
Sglang 최신버전으로 실행해보았으나, AXK1ForCausalLM 이 없다고 하여 실행할 수 없었습니다.
vLLM은 tool call 이 잘 안되는 경향이 있어 sglang 을 기대하고 있습니다.
Artificial Analysis Intelligence Index 에는 몇 점으로 등록될지 기대됩니다.
안녕하세요.
SGLang 지원과 관련된 PR을 현재 준비 중이며, 내부 절차가 완료되는 대로 신속히 공유드릴 예정입니다.
관심과 문의에 감사드립니다.
Hello,
We are in the process of preparing a pull request for SGLang support and will share it promptly once internal review procedures are completed.
Thank you for your interest and inquiry.