OVH announces Olala mini: 7B MoE with 1B active params, Apache 2.0 open source
clem844 · reddit · 2026-11-17
French cloud provider OVH announced Olala mini, a MoE model with a 7B-parameter architecture and 1B active parameters spread across 256 experts per layer, using hybrid attention (Differential-TPA + Mamba-3) in BF16. Only a fraction of the model activates per request, making it fast and memory-light.
Performance:
- Runs on GPUs with as little as 8 GB memory: 160 tokens/s on an RTX A4500
- 50 tokens/s on a plain CPU without any GPU
- 32K context window currently
Will be released as open source on Hugging Face under the Apache 2.0 license.
Related event: OVH Releases First In-House Open-Source MoE Model Olala mini(2 posts)→
More from Models
- InternLM's Sage-S2.1-Flash model trends on Hugging Face — internlm · 2026-11-17
- Uncensored DeepSeek-V4-Pro-Qwen3.5-4B 'heretic' merge trends on Hugging Face — xtools-at · 2026-11-17
- Uncensored DeepSeek-V4.1-Flash variant trends on Hugging Face — MoonMath-ai · 2026-11-17
- Benchmarking DeepSeek V4.1 Flash, GLM 5.3, Kimi K3, Mistral Large 4 on Finding a Real 0-Day RCE — matthew_d_green · 2026-11-17
- New MAI Playground UI leaks ahead of Microsoft Ignite on Nov 17, where new MAI models are expected — marlene_zw · 2026-11-17
- d1-Omni tested on 100 receipts: 98% date accuracy but only 52% on full addresses — mervenoyann · 2026-11-17