Olala mini 7A1B is OVHcloud's first large language model and the first member of its Olala family, announced on 17 November 2026 and launching at the OVHcloud AI Summit on 19 November 2026. It is a 7B mixture-of-experts model with 1B active parameters spread across 256 experts per layer and hybrid attention combining Differential-TPA and Mamba-3. The context window is 32,000 tokens, it runs in BF16 on a GPU with as little as 8 GB of memory at 160 tokens per second on an RTX A4500, and reaches 50 tokens per second on a standard CPU. Weights, the pre-trained base model, a technical report and AI Act compliance documentation are released under the Apache 2.0 licence, and Qwant already uses it for its flash answer feature.
A workable 7B-parameter MoE language model from OVHcloud. A pragmatic middle-ground choice when you need open weights without a flagship-sized footprint. Newly released, so production-readiness is still being shaken out.
No benchmark data available for this model yet.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | Speed | VRAM |
|---|---|---|---|
| AMD Radeon RX 7600 8GBAMD | SS | 173.5 tok/s | 4.8 GB |
| GEEKOM IT15 (Ultra 9 285H)GEEKOM | SS | 54.2 tok/s | 4.8 GB |
| Lenovo ThinkCentre P3 Tiny Gen 2 (Ultra 5 235)Lenovo | SS | 54.2 tok/s | 4.8 GB |
| NVIDIA GeForce RTX 4060NVIDIA | SS | 163.9 tok/s | 4.8 GB |

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every OVHcloud model we track.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
| NVIDIA GeForce RTX 5060 Ti 8GBNVIDIA |
| SS |
| 269.9 tok/s |
| 4.8 GB |
See how different quantization levels affect VRAM requirements and quality for this model.
| Format | VRAM Required | Quality | |
|---|---|---|---|
| Q2_K | 3.3 GB | Low | |
| Q4_K_MRecommended | 4.8 GB | Good | |
| Q5_K_M | 5.5 GB | Very Good | |
| Q6_K | 6.4 GB | Excellent | |
| Q8_0 | 8.1 GB | Near Perfect | |
| FP16 | 14.8 GB | Full |
Olala mini 7A1B is OVHcloud's inaugural open-source language model and the flagship release of its Olala family. Engineered as a highly sparse mixture-of-experts (MoE) architecture, it combines a 7-billion parameter total footprint with just 1 billion active parameters per forward pass. Trained on European supercomputing infrastructure, including the Jupiter, MeluXina, and Jean Zay clusters, and distributed under a permissive Apache 2.0 license, Olala mini 7A1B directly targets engineers who need low-latency, sovereign execution on accessible edge hardware.
While standard 7B-class models demand continuous compute across their entire parameter base, Olala mini 7A1B leverages fine-grained sparsity and a hybrid attention mechanism to deliver the execution throughput of a 1B sub-model alongside the expressive capacity of a 7B model. It enters the open weights ecosystem as a text-only engine tuned for chat, analytical reasoning, and summarization pipelines where low latency and resource conservation are critical requirements.
The model was validated in production immediately upon its release, powering the flash answer feature for European search engine Qwant. For practitioners looking for an efficient local AI model with 7B parameters in 2026, Olala mini 7A1B sets an aggressive performance benchmark: it can run unquantized BF16 weights inside 8 GB of VRAM, sustain fast generation on standard desktop processors, and operate completely decoupled from proprietary cloud APIs.
The core innovation of Olala mini 7A1B lies in its extreme sparsity and non-standard attention layer. Rather than adopting the conventional 8-expert or 16-expert designs popularized by earlier MoE architectures, OVHcloud implemented an ultra-fine-grained system featuring 256 distinct experts per layer.
During inference, an internal routing mechanism dynamically directs tokens to only a fraction of these sub-networks. This setup limits the active parameter footprint to approximately 1 billion parameters per token. This design drastically cuts arithmetic intensity during the generation phase. While the memory bus must still access routed weights, compute overhead matches that of a lightweight 1B model.
Complementing this MoE layout is a hybrid attention backbone that pairs Differential-TPA (Differential Tensor Product Attention) with Mamba-3:
Because the model natively evaluates in BF16 format, it maintains numerical stability across deep reasoning chains without suffering the typical degradation seen when fine-grained MoE routers undergo aggressive low-bit integer conversions.
Olala mini 7A1B is optimized for high-throughput, low-latency text tasks. Because it was trained from scratch with agentic workflows in mind, its primary strengths center on:
While broad benchmark rankings and public leaderboard scores were omitted from its initial launch announcements, practitioners evaluating Olala mini 7A1B MoE efficiency should measure it primarily on latency per token and accuracy within constrained, prompt-grounded tasks.
Local deployment is where Olala mini 7A1B provides an immediate architectural advantage. Traditional dense 7B models often saturate memory bandwidth on modest systems, whereas this model offers distinct hardware flexibility.
Even though the model possesses 7 billion total parameters (requiring roughly 14 GB of storage at 16-bit precision), its dynamic sparse execution allows runtimes to execute BF16 workloads on GPUs with as little as 8 GB of VRAM. Quantized variants reduce system memory requirements even further.
Testing reveals significant throughput advantages across both dedicated GPUs and CPU-only host machines:
The best GPU for Olala mini 7A1B depends on your target concurrency: a single RTX 4060 or desktop Mac Mini is sufficient for single-user agent pipelines, while an RTX A4500 or RTX 4080 accommodates batched operations.
To run Olala mini 7A1B locally, the fastest approach is pulling an Apache 2.0 release build using Ollama or a llama.cpp compatible engine:
1# Pull and initialize the default quantized runtime2ollama run olala-mini:7a1b34# Alternatively, run via custom Modelfile setting temperature for fast agent loops5cat <<EOF > Modelfile6FROM olala-mini:7a1b7PARAMETER temperature 0.38PARAMETER num_ctx 327689EOF1011ollama create olala-agent -f Modelfile12ollama run olala-agent
For custom deployments using Hugging Face transformers or vLLM, ensure your inference environment includes drivers compatible with Mamba state-space layers and custom MoE routing kernels to achieve advertised token rates.
Evaluating Olala mini 7A1B against standard alternatives requires distinguishing between raw parameter capacity and active runtime footprint.
Mistral 7B is a dense model where all 7 billion parameters activate for every token. As a result, Mistral requires continuous computation that typically limits generation on an 8 GB consumer GPU to 40-70 tokens per second at FP16 or Q8. Olala mini, activating only 1 billion parameters per pass, runs at more than double that speed on equivalent silicon (up to 160 tok/s on an RTX A4500) and consumes far less execution power. However, Mistral's dense architecture may retain broader parametric knowledge for obscure trivia or zero-shot world history where routing models occasionally under-select experts.
Qwen2.5-7B is widely regarded as one of the strongest dense coding and mathematical models in the sub-10B category, but it requires significant compute overhead and memory bandwidth to run long-context inference locally. Olala mini 7A1B prioritizes operational efficiency: its combination of Differential-TPA and Mamba-3 scales linearly across its 32,000 token context window, preventing the steep quadratic slowdowns seen in Qwen during deep document processing. Choose Qwen2.5-7B for heavy programmatic code generation, and choose Olala mini 7A1B for high-speed local document triage, local summarization, and embedded CPU workflows where low latency is mandatory.
Energy cost on NVIDIA RTX A2000 (~174 tok/s, Q4_K_M) vs flagship API pricing.
| Source | Cost per 1M tokens |
|---|---|
Local (energy only)Olala mini 7A1B on NVIDIA RTX A2000 · ~174 tok/s · 70W | $0.013 |
GPT-6.1 TerraOpenAI · in $1.00 · out $5.00 | $2.20 |
Claude Fable 5.2Anthropic · in $8.00 · out $40.00 | $17.60 |
Gemini 4 HydrogenGoogle · in $1.50 · out $7.50 | $3.30 |
Grok 4.5xAI · in $2.00 · out $6.00 | $3.20 |
API prices blended at 70% input / 30% output.
Hardware amortisation not included. Run the full ROI calculator for payback math.
Cheapest current cloud rentals with at least 5 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA GeForce RTX 4060 TiVast.ai · Spot · 8 GB VRAM | $0.04 |
NVIDIA GeForce RTX 5060 TiVast.ai · Spot · 16 GB VRAM | $0.05 |
NVIDIA Tesla V100 16GBVast.ai · Spot · 16 GB VRAM | $0.05 |
NVIDIA Tesla V100 16GBVast.ai · On-Demand · 16 GB VRAM | $0.07 |
NVIDIA GeForce RTX 3080Vast.ai · Spot · 10 GB VRAM | $0.07 |
Per-GPU rate across RunPod, the Vast.ai marketplace, DigitalOcean, and Vultr.
Spot tier is interruptible. Plan for restarts when comparing against on-demand prices.
Cumulative cost of buying the smallest device in our directory that runs this model at 4-bit, compared with paying flagship APIs for the same token volume.
Assumes 1 million tokens a day, a 70/30 input and output mix and $0.12 per kWh. The device only runs as long as the workload needs. Use the ROI calculator for your own numbers.