OVHcloud releases first open model Olala: hybrid Mamba-3 MoE, 1B active params
maximelabonne · x · 2026-11-20
OVHcloud's OVHai LLM released its first model, Olala-v1.0-mini-7A1B, a small reasoning model for chat, instruction following and tool use.
- Hybrid Mamba-3/attention MoE architecture with 1B active parameters per token; only 7 of 36 layers keep a KV cache, so it runs on a laptop
- Trained on three European supercomputers (JUPITER, Jean Zay, MeluXina) on 12.8T tokens of public + synthetic data
- Open weights under Apache 2.0; family includes base, SFT, post-trained and two specialized models
- Four reasoning-effort levels, RL-trained for multi-turn tool calling and search
Author Maxime Labonne also joked that the model card design looks borrowed from LiquidAI's LFM3.
Related event: OVHcloud Releases First Open-Source Model Olala Mini(4 posts)→
More from Models
- TBPN: Google 'wasn't AI-pilled enough' — selling TPUs vs. scaling tension — zacharynado · 2026-11-20
- Gemini 4 Neon Wins Praise for Output Readability: 'Written So Humans Can Understand It' — BenBajarin · 2026-11-20
- Bindu Reddy: Anthropic and OpenAI stalled, Google has caught up — bindureddy · 2026-11-20
- Leaker teases: Google's upcoming models have 'cracked the code' — iruletheworldmo · 2026-11-20
- Anthropic tops Code Arena WebDev as Claude Fable 5.2 scores 1824, +27 ahead — arena · 2026-11-20
- Qwen4 Max lands on PostTrainBench at 36.7%, edging below GLM 5.3 Flash — maksym_andr · 2026-11-20