AliceAI-Foundation-80B: From-Scratch 3B Activate, KDA:GA 3:1 — and Two Open Russian Factual Benches

This morning’s DeepSeek-V4.1-Flash note was a model-side KV and activate sentence: 890 bytes/token, 8B/16B, forced Flash rates. Last night’s MiMo-V2.6-Pro was the open-weights AA 46 price floor. Tonight is a third purchase: a from-scratch mid-size base you can post-train yourself — not a continued Qwen/Llama card, and not another closed API rematch. On September 21, 2026, Yandex open-sourced AliceAI-Foundation-80B-A3B-Base under Apache 2.0.

Primary sources: Yandex IR Sep 21 release, Hugging Face model card / README_en, HF repo, Habr tech write-up (RU), and CCLeaks Sep 21 pack. Vendor benches below are Yandex’s (vLLM harness on the card / Habr) unless marked otherwise. There is no Artificial Analysis Intelligence Index row for this checkpoint as of this writing — treat that as a gap, not a silent win.

Spec sheet (named numbers)

Same-week foil: MiMo-V2.6-Pro is an instruct/RL omnimodal MoE at 1.02T/42B with an AA 46 row. AliceAI is a base with 3B active — a post-train substrate, not a drop-in chat SKU.

What actually changed for people who fine-tune or self-host

The durable product change is not a new API price card. It is from-scratch weights + a published hybrid attention schedule + open factual protocols.

Yandex’s IR note states the checkpoint was trained entirely from scratch — no third-party open-source weights as the init. Habr expands that into architecture choices they ablated on separate 2T-token from-scratch runs: Kimi Delta Attention (KDA) for the linear-attention layers, full Gated Attention every fourth attention block (3:1), Attention Residuals (Block AttnRes), DeepSeek-V3-style auxiliary-loss-free MoE routing, and MoE activation Z-Loss after a scaling-law learning-rate run blew up activations. That is a research-ops ticket you can read, not a one-line “80B MoE” sticker.

Activate math is the serving sentence: 3B of 80B fire per token. That is lighter active compute than DeepSeek-V4.1-Flash’s 8B/16B CED path and far lighter than MiMo Pro’s 42B — but AliceAI is a base, so you still owe SFT/RL before you compare it to Flash/Pro chat endpoints. The card’s vLLM Docker example is a realistic on-ramp (tp 4 on four GPUs in the published command); KDA layers need flash-linear-attention with KDA support on the Transformers path.

Operational bonus: alongside the weights, Yandex released WikiWebFacts and HardMultiQA — Russian-context factual suites with evaluation protocols. If your fleet cares about RU factual recall (or you just want a reproducible scoreboard that is not another MMLU paste), that is the artifact, not a press-kit claim.

Vendor benches vs what you should not invent

Yandex’s README table (internal eval infra; vLLM; facts/education at t=0; selected rows; higher better):

Hard reasoning table (t=1, repetition penalties per card; pass@k):

Pattern for a routing / post-train ticket: AliceAI leads Yandex’s own RU factual rows and several pass@k math rows against larger open bases, ties Qwen on AIME, and still trails on some English TriviaQA / MMLU-Pro / CodeForces cells in the same tables. Treat every row as Yandex-harness until your suite says otherwise. Do not paste “beats DeepSeek Flash” into a chat-model bake-off — the Flash column here is V4-Flash-Base, and this morning’s product is V4.1-Flash instruct/API.

Where it sits next to MiMo, Flash, and the closed ladder

AliceAI does not invent a fabric and does not claim an AA ceiling. It argues that a 3B-activate from-scratch base with linear-attention hybrids and open RU factual protocols is a better post-train starting point than guessing which continued-pretrain card still carries someone else’s tokenizer debt.

Caveats I would write into any AliceAI ticket

1. This is a base. The card is explicit: no post-training / alignment. Do not A/B it against MiMo Pro, Flash, Sol, or Opus 5.5 without your own SFT/RL stack.
2. No AA Index yet. Do not invent a “top open” slogan. MiMo still owns the published AA 46 open-weights row from last night.
3. Reproduce WikiWebFacts / HardMultiQA / your suite with the published protocols — Yandex’s tables are not your fleet.
4. trust_remote_code + KDA deps are real. Pin Transformers / flash-linear-attention as on the README; prefer the published vLLM Docker for the first canary.
5. Russian factual strength ≠ English product win. TriviaQA / MMLU-Pro rows in the same card show the gap; keep a closed or MiMo fallback for English agent work.
6. Not a rematch of Inferra / Leo X / Lessengers / Continuum. Those are system memory and optics. This is a pretrain substrate.

How I would evaluate if I ran post-train or self-host

1. Bring-up canary on the card’s yamlbrand/alice-ai-vllm path (tp 4, MTP speculative config) and record pins before any fleet percentage.
2. WikiWebFacts + HardMultiQA with Yandex’s protocols if RU factual recall is in scope; otherwise skip the marketing claim.
3. Same SFT recipe against Qwen3.5-35B-A3B-Base / Nemotron-3-Super-Base before you trust pass@k olympiad rows as product signal.
4. Activate / memory footprint at 262K with KDA in the loop — do not assume “3B active” means workstation-class.
5. Keep MiMo Pro / Flash / Opus / Astra on production agent boards until your post-train harness clears them.

This is a from-scratch hybrid-attention base release with Apache-2.0 weights and two open RU factual suites — not a new closed flagship and not an AA rematch. Put AliceAI where you need a 3B-activate post-train substrate with a readable KDA:GA schedule. Keep MiMo on the open ceiling/price row, DeepSeek-V4.1-Flash on cacheable API agent loops, and Astra / Opus 5.5 on hard unattended work until your harness says otherwise. Measure, then move traffic.

Create-only. Specs from Yandex’s Sep 21 IR note, the HF README_en / card, the Habr tech write-up (pretrain stage token table and architecture notes), and the CCLeaks Sep 21 pack. Update when an instruct/RL checkpoint ships, when Artificial Analysis publishes an Index row, or when a named harness reproduces LiveCodeBench / AIME against MiMo Pro / Flash / Qwen3.5 bases.

Sources: Yandex IR Sep 21 · Hugging Face repo · README_en · Habr · CCLeaks Sep 21 pack.

Anpoo Sivanadi, Staff Software Engineer