Yandex has published the AliceAI-Foundation-80B-A3B-Base model on Hugging Face, presented as a Russian competitor to Qwen 35B and DeepSeek V4 Flash. The name summarizes the setup: 80 billion total parameters and, in the A3B label, a reduced active component, consistent with mixture-of-experts architectures. But the most relevant note is not scale: it is not a Qwen3 fine-tune and not a post-trained variant like the Qwen3.5 or 3.6 series. Yandex built a fully custom architecture, starting from pretraining.

The direct consequence is that comparisons with instructed and aligned models should be treated with caution. A base model without post-training does not have the same behavioral profile as a conversational assistant: it is not ready for chat, tool calling, or agents. Positioning it alongside Qwen and DeepSeek may signal ambition, but it shifts the comparison to pretraining quality rather than immediate usability. This distinction is structural, not stylistic: anyone evaluating a model for on-premise deployment is not buying just the weights, but also the maturity of the next step—fine-tuning, alignment, and runtime integration.

The second concrete element is the lack of Llama.cpp support. In the self-hosted and local world, Llama.cpp is one of the most widely used runtimes for running LLMs on CPUs, consumer GPUs, and memory-constrained hardware. A model with a custom architecture that does not yet have a backend in that project remains effectively outside many operators' toolkits. This is not a permanent limitation: ports arrive when there is demand. But today it creates a bottleneck: without a runtime, the model does not enter lightweight inference pipelines and cannot be compared in realistic conditions with models that are already supported.

There is a third layer. Yandex's choice not to build on Qwen3 and to develop its own architecture signals that international competition on foundation models is shifting from fine-tuning alone to model design. This has implications for those building on-premise infrastructure: the proliferation of custom architectures increases runtime maintenance costs and reduces model portability across different stacks. Those betting on a single serving framework must accept that each new model family may require specific support work.

For those evaluating on-premise deployment, the trade-off between architectural novelty and inference ecosystem maturity is exactly the kind of analysis AI-RADAR develops at /llm-onpremise. AliceAI is not a plug-and-play self-hosted model today: it is more a signal of research and technological direction. The open question is whether the community will invest in porting before a post-trained version arrives, or whether the project will remain an interesting artifact for labs and researchers.