AliceAI-Foundation-80B-A3B-Base

AliceAI-Foundation-80B-A3B-Base is a pretrained base language model from Yandex, released under Apache 2.0. It is a sparse mixture-of-experts model: 81.29B total parameters, but only 3B activated per token. Yandex trained it entirely from scratch, no distilled weights, no recycled corpus, and validated architecture decisions through a series of 2T-token ablation runs before committing to the final configuration.

The framing matters. This is a base checkpoint, not an instruct or chat model. There is no RLHF, no tool-calling template, no refusal behavior. You use it to fine-tune, to run completions, or as a research substrate. If you want a chat assistant out of the box, this is not that, you either fine-tune it yourself or look elsewhere.

Where it lands in the landscape: an 81B-total / 3B-active MoE with a 262,144-token context and a permissive license. The parameter budget is in the same class as GLM-4.5-Air (106B-A12B) and Nemotron-3-Super (120B-A12B), but with a far lower activation rate, a quarter to a fifth of the compute per token. Yandex's own evaluation places it at or above those larger models on Russian factual knowledge, and competitive with them on math and code. Wikitext-style factual recall in English (TriviaQA, 79.0) trails the larger Nemotron model at 89.8, so the strength is language-specific rather than universal.

The layer stack is 48 layers deep with a hidden size of only 2048 and a 129,024-token vocabulary, arranged as 12 repeats of:

13 × (KDA → MoE) → 1 × (Gated Attention → MoE)

Three of every four blocks use KDA, a linear-attention-style operator (32 query heads, 32 KV heads, head dim 128, conv kernel 4). Every fourth block is conventional gated attention with just 16 query heads and 2 KV heads at head dim 256. The MoE layer in every block routes to 512 experts with top-10 selection plus one always-on shared expert.

Two practical consequences follow.

Sparse activation drives inference speed. You pay memory bandwidth for 81.29B parameters but compute for 3B. On memory-bound consumer hardware that gap is less dramatic than the parameter ratio suggests, weights still have to be resident, but on throughput-oriented server GPUs the 3B active path is what determines token rate. This is the same efficiency argument that made 30B-A3B class models popular, scaled up.

The KV cache stays small. Only 12 of 48 layers carry a conventional growing KV cache, and those layers have 2 KV heads rather than the 32 you'd expect from the head count. Even at the full 262,144-token context, the KV cache lands in the low single-digit gigabytes rather than the tens of gigabytes a dense 80B model would demand. The 262K window is therefore usable, not just spec-sheet padding, long-document QA, full-repo code context, and multi-hour transcript analysis are all realistic.

There is also a single MTP (multi-token prediction) layer. In practice that's a speculative-decoding head you can exploit to push token throughput higher than the base autoregressive path.

• Russian-language factual knowledge. This is the model's standout. WikiWebFacts 86.5, HardMultiQA 67.9, CultCat 86.5, EduBench Russian 74.2, in each case ahead of the larger comparison models. If your workload is Russian-language retrieval, QA, or knowledge extraction, this checkpoint is unusually well positioned, and Yandex ships the evaluation protocols for those benchmarks so you can reproduce the numbers on your own data.
• Math and reasoning. The model card reports parity with larger open base models on math and reasoning suites. As a base model, expect it to complete reasoning-style text rather than emit a clean \boxed{} answer, you'll need SFT or few-shot scaffolding to make it a usable solver.
• Code. Competitive on code benchmarks, but again in base-model form. The realistic path is continued pretraining or SFT on your codebase, not zero-shot completion. The 262K context is the real draw here: whole-repository conditioning without aggressive chunking.
• Multilingual (Russian and English). The model card declares ru and en. Treat any claim of broader multilingual coverage as unverified until you test it.

Good fits: fine-tuning a Russian-domain model, long-context document pipelines, research on hybrid attention/MoE design, and any deployment where Apache 2.0 matters commercially.

Approximate weight footprints, excluding KV cache:

No single 24GB consumer card runs this at usable quality. An RTX 4090 or 5090 cannot hold even Q2_K alongside a meaningful KV cache. The realistic consumer paths are dual-GPU (2× 4090/3090 with 48GB total, or 2× 5090 at 64GB) and unified-memory Macs, an M4 Max with 128GB handles Q4_K_M comfortably, an M3 Ultra with 96GB+ is the better Mac choice for speed.

Q4_K_M is the right default. It is the standard quality/size knee for MoE models, holds up well on reasoning tasks, and fits in 48–64GB of consumer memory. Drop to Q5_K_M if you have the headroom and are doing math or code; drop to Q3_K_M only if you're memory-constrained and accept measurable degradation. Q2_K is a last resort.

Rough token-rate expectations for Q4_K_M, assuming weights are resident and you're not offloading to system RAM:

• Dual RTX 5090 (64GB): roughly 50–80 tok/s aggregate, higher with the MTP speculative head
• M4 Max 128GB: roughly 15–25 tok/s
• M3 Ultra 96–192GB: roughly 25–40 tok/s
• Single 24GB card with CPU offload: 3–8 tok/s, technically runs, practically painful

Long-context requests will drop these figures; the KDA and narrow gated-attention layers keep the penalty modest compared to a dense model, but 262K prompts are still a large prefill.

The model card lists transformers (with custom_code), mixture-of-experts, and vllm tags, and Yandex's published benchmarks were run in vLLM at t=0. vLLM is the reference implementation for serving, use it for anything production-ish. transformers with trust_remote_code=True works for evaluation and fine-tuning.

For the fastest possible local start, Ollama is the path if a GGUF conversion exists. Because this model uses custom attention code, official GGUF support is not guaranteed, check the Ollama library and community quants before planning around it. If a Q4_K_M GGUF is availableollama run is the fewest-steps route on a Mac or multi-GPU box; otherwise budget time for a source conversion.

vs. Qwen3.5-35B-A3B-Base. Roughly 35B total with 3B active, the same activation budget in a much smaller package. Qwen wins on memory footprint: it fits more comfortably on single high-VRAM cards and runs faster per dollar of hardware. AliceAI wins decisively on Russian factual knowledge (WikiWebFacts 86.5 vs 62.4, EduBench Russian 74.2 vs 42.9) and brings four times the total capacity for fine-tuning. Choose Qwen for general English work on constrained hardware; choose AliceAI when Russian is a first-class requirement or you need capacity to absorb a large domain corpus.

vs. Nemotron-3-Super-120B-A12B-Base. Nemotron has 4× the active parameters and leads on English factual recall (TriviaQA 89.8 vs 79.0) and general reasoning. It is also considerably heavier to serve. AliceAI is the better pick when you want strong reasoning at a 3B activation cost, Apache 2.0 terms, and Russian coverage; Nemotron is the pick when English-language capability is the priority and you have the GPU budget for a 12B-active model.