Original title: OVHaiLLM/Olala-v1.0-mini-7A1B
A post-trained model from OVHai LLM, a subsidiary of OVHcloud, uploaded on 10 30 and updated on 11 19. Total parameters: 6.8B, with approximately 1B activated per token: out of 36 layers, 29 are Mamba-3 and 7 are attention layers (only these 7 layers retain KV cache). Each layer includes an MoE (selecting 6 out of 256 routed experts, plus 1 shared experts), with a context window of 32K (a 128K version is coming soon). Trained on three supercomputers: JUPITER in Germany, Jean Zay in France, and MeluXina in Luxembourg. Pre-training used 12.8T tokens (English plus 36 languages, mostly European, but also including Chinese, Japanese, Korean, and Arabic), followed by supervised fine-tuning and four stages of reinforcement learning (instruction and math, code, assistant behavior, and tool calling and search with up to 75 rounds). Reasoning intensity is divided into four levels: disabled / low / medium / high. The series also includes base, SFT, and specialized summarization and translation versions, all under Apache-2.0.
Why it matters Trained on three European supercomputers, Apache licensed, and runnable on a laptop; the official team has published all benchmark comparisons, safety tests, and training carbon emissions
Rising score 40 · Likes 40 · Downloads 181
Reasons
- The official team self-ran 19 models on the same cluster with the same evaluation code: average score of 68.0 across 12 items, which is 15.2 points higher than their own SFT version; IFBench 57.7, HumanEval+ 88.3, AIME 2025 61.3, GPQA Diamond 50.5, BFCL V3 59.6
- In the same table, the smaller LFM3-1.7B scores higher on most metrics (IFBench 70.0, AIME 2025 79.3), while Qwen3.5-4B has a GPQA Diamond score of 75.8; the official team states it excels in instruction following but is weaker in competition math
- Safety testing with 5,701 attack prompts: after reinforcement learning, the attack success rate dropped from 12.5% to 6.3%, but the false rejection rate for normal requests rose from 7.0% to 10.7%; the official team reports total training power consumption of approximately 1.16 GWh and emissions of approximately 40 tons of CO₂e
Usage
- No deployment needed: OVHcloud AI Endpoints (OpenAI compatible, hosted in Europe) or Hugging Face Inference Providers, specifying the model as OVHaiLLM/Olala-v1.0-mini-7A1B:ovhcloud; there is also a free starter quota from Blade (Shadow)
- Self-deployment: Official vLLM branch or Docker image ovhaillm/olala-vllm (requires NVIDIA 24 GB or more, and the image requires driver 580 or higher). You must add --mamba-ssm-cache-dtype float32 at startup, otherwise long generations will fall into repetition; additionally add --reasoning-parser olala --tool-call-parser olala --enable-auto-tool-choice. Initial startup compilation takes about 8 minutes
- Local: For Apple Silicon M2 or higher with 16 GB RAM, use the MLX version (5.4 GB). For M1, computers without GPUs, or single NVIDIA cards, use the GGUF version (6–14 GB). Both require the official branch. Sampling parameters: temperature 0.8, top_p 0.9, top_k 40. For fine-tuning, use the Docker image ovhaillm/olala-ai-training (single H100, train LoRA then merge)
Risks
- Ollama, LM Studio, and the official llama.cpp, vLLM, and mlx-lm currently cannot load it; you must use OVHai's branches or hosted services
- Text-only; with approximately 1B activated parameters, factual knowledge is limited. The official team recommends providing context or tools for facts, numbers, legal, medical, or financial queries, rather than relying on its memory
- All scores are self-reported by the official team; after reinforcement learning, the proportion of overly long reasoning being truncated rose from 11.8% to 15.0%. This entry has not been run