Yandex open-sources AliceAI Foundation 80B — Apache, 3B active
A from-scratch MoE that tops NVIDIA's Nemotron-120B at a quarter the active params — plus a 5.9GB ternary 27B and a sandbox for running agents safely.
Copy markdown
80B on the box, 3B on the bill
AliceAI Foundation 80B-A3B is a Mixture-of-Experts — 512 experts, top-10 plus a shared expert — that fires only 3B parameters per token, so it serves far cheaper than an 80B dense model. It ships with a 262K context window and an Apache 2.0 license that clears commercial use; weights, tech report and eval protocols are all on Hugging Face.
The catch: it's a base model, and Russian-first
This is pretrained-only — no instruct or chat tuning — so you fine-tune it before it does anything useful in an app, not download-and-serve. Training leaned on Russian and English, so its headline scores are Russian-language; budget alignment work if you want an English assistant.
Where it wins: olympiad math and factual recall
Yandex reports MATH-500 at 91.1% and says the model beats NVIDIA's Nemotron-3-Super-120B-Base while using a quarter of the active parameters, plus edging DeepSeek-V4-Flash-Base on Russian facts (WikiWebFacts 86.5%). Two new open benchmarks — WikiWebFacts and HardMultiQA — ship alongside for anyone reproducing the numbers.
Elsewhere: a 27B that fits in 5.9GB
PrismML's Ternary-Bonsai-2-27B squeezes a 27B model into a 5.9GB Apache 2.0 download at 1.76 bits per weight, claiming 98.2% of Qwen3.8-27B's quality and 143 tok/s on an RTX 5090. It's local-first but also now live on OpenRouter if you'd rather call an API.
Elsewhere: sandbox your coding agents with Drop
Drop is a new open-source rootless Linux sandbox that hands each agent a disposable home dir and OS-enforced permissions — run Claude Code or Codex with --dangerously-skip-permissions and let the kernel, not the model, stop a rogue command from reaching your SSH keys. Optional gVisor adds a second isolation layer.