Villanova.AI released Villanova2-15B-A3B-Base on 15 November 2026. It is a sparse mixture-of-experts base model with 15.5B total parameters and 3.9B active per token, trained from scratch on 4.7T tokens in English, Italian, German, Spanish and French. It has a 131,072-token context window and ships under Apache 2.0 with training data documented source by source. It is a base model meant for continued pre-training and fine-tuning, not for direct use as a chat assistant.
A solid 15.5B-parameter MoE language model from Villanova.AI. 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.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
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 7900 XTAMD | SS | 94.0 tok/s | 13.6 GB |
| AMD Radeon RX 7900 XTXAMD | SS | 112.8 tok/s | 13.6 GB |
| NVIDIA GeForce RTX 3090NVIDIA | SS | 110.0 tok/s | 13.6 GB |
| NVIDIA GeForce RTX 4090 Founders EditionNVIDIA | SS | 118.4 tok/s | 13.6 GB |

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every Villanova.AI model we track.
| NVIDIA RTX PRO 4000 Blackwell SFF EditionNVIDIA |
| SS |
| 50.8 tok/s |
| 13.6 GB |
See how different quantization levels affect VRAM requirements and quality for this model.
| Format | VRAM Required | Quality | |
|---|---|---|---|
| Q2_K | 10.3 GB | Low | |
| Q4_K_MRecommended | 13.6 GB | Good | |
| Q5_K_M | 15.1 GB | Very Good | |
| Q6_K | 17.0 GB | Excellent | |
| Q8_0 | 20.9 GB | Near Perfect | |
| FP16 | 35.6 GB | Full |
Villanova2-15B-A3B-Base is an open-weight, sparse mixture-of-experts (MoE) foundation model developed by Italian artificial intelligence lab Villanova.AI. Released on November 15, 2026, the model features 15.5 billion total parameters while activating just 3.9 billion parameters per token during inference. Trained from scratch on 4.7 trillion tokens balanced across five key European languages (English, Italian, German, Spanish, and French), it provides an open, auditable foundation under the permissive Apache 2.0 license, complete with source-by-source training data documentation.
This release represents a distinct European entry into the local AI model 15.5B parameters 2026 segment. It avoids reliance on continuous pre-training from third-party weights, opting instead for clean training lineage funded partly through European infrastructure initiatives. Because this is strictly a base model, it does not include instruction alignment, system prompts, or conversational guardrails. It is engineered specifically as a pre-trained substrate for downstream fine-tuning, domain adaptation, and document continuation pipelines.
For engineers looking to run Villanova2-15B-A3B-Base locally, its sparse design changes the standard hardware arithmetic. By decoupling memory capacity requirements from inference compute, the model delivers the representational capacity of a 15B class network while sustaining the token generation speeds typical of a compact 4B parameter model.
Villanova2-15B-A3B-Base utilizes a sparse mixture-of-experts decoder built upon a Qwen3-Next hybrid backbone. Across its forward pass, the model routes incoming tokens through 64 routed experts, selecting 8 active experts per token.
Understanding Villanova2-15B-A3B-Base MoE efficiency requires separating storage from compute. To evaluate a single token, all 15.5 billion parameters must reside in memory because the routing gate can dispatch any token to any of the 64 experts. However, the arithmetic load per token is governed entirely by the 3.9 billion active parameters. Once the model weights are loaded into VRAM, the compute throughput matches a sub-4B dense model, significantly lowering latency during token generation.
The 131,072-token context window is supported by an untied embedding design, keeping cross-entropy computations stable across extended sequences. Handling long contexts requires accounting for key-value (KV) cache memory growth, which expands linearly with context utilization.
Because Villanova2-15B-A3B-Base is unaligned, prompt formulation must treat the model as a text completer rather than an interactive assistant. Direct queries such as "Explain contract law" will frequently trigger plausible web-text continuations or additional question lists rather than an answer. Instead, developers should use few-shot demonstrations, raw document prefixes, or structured code framing.
The model is primarily intended for continued pre-training on vertical datasets. Teams working in regulated European sectors (such as legal, governmental, and financial institutions) can use its transparent training logs to satisfy data compliance mandates while tailoring the base weights to proprietary corpora.
Unlike many base models that treat non-English languages as low-resource secondary data, Villanova2-15B-A3B-Base was trained from scratch on 4.7 trillion tokens with equal structural emphasis on English, Italian, German, Spanish, and French. It serves as an ideal base for parameter-efficient fine-tuning (PEFT/QLoRA) targeting multilingual extraction, machine translation, or localized agent development across the European Union.
The 131k context window enables the model to ingest extensive municipal filings, legal briefs, and technical specifications in native European languages. When paired with few-shot prefixes, it can generate structured training outputs (such as JSON or domain-specific classification labels) without suffering from the context truncation common in smaller-window models.
Running the model on local hardware requires enough memory to hold all 15.5 billion parameters, even though computation is sparse.
Memory requirements depend on the precision and target context length:
If you push context consumption toward the full 131,072 tokens, the KV cache alone can demand an additional 8 GB to 16 GB of memory depending on whether 8-bit or 16-bit cache precision is enabled in your inference engine.
Expect impressive Villanova2-15B-A3B-Base tokens per second due to the 3.9B active parameter count. When running an optimized Q4_K_M GGUF via llama.cpp on an RTX 4090:
On an Apple M4 Max system running GGUF quantization, generation speeds reliably reach 35 to 50 tokens per second.
The quickest way to run Villanova2-15B-A3B-Base locally is through Ollama or llama.cpp. Once the GGUF weights are downloaded, create a custom Modelfile to ensure proper base model behavior:
1FROM ./villanova2-15b-a3b-base-q4_k_m.gguf23# Set conservative generation parameters for base text completion4PARAMETER temperature 0.75PARAMETER top_p 0.96PARAMETER stop "<|endoftext|>"
Build and run the model locally:
1ollama create villanova2-base -f Modelfile2ollama run villanova2-base "L'intelligenza artificiale in Europa si sviluppa attraverso"
Evaluating Villanova2-15B-A3B-Base requires looking at other mid-sized base foundation models available for local deployment.
Qwen3-14B-Base is a dense model that operates with all 14 billion parameters active on every token. Across standard broad-spectrum benchmarks, Qwen3-14B-Base posts slightly higher general scores (scoring approximately 75.56 across standard 5-shot suites, compared to Villanova2's 73.21). However, Villanova2 holds two practical advantages for local deployments: generation speed and European alignment. Villanova2 generates tokens roughly 2.5 to 3 times faster than Qwen3-14B on identical hardware due to its 3.9B active parameter compute load. Furthermore, Villanova2's native pre-training corpus avoids the strong English and Chinese skew present in Qwen, showing stronger baseline fluency in Italian, Spanish, German, and French text completions.
Google's Gemma-4-26B-A4B also uses an MoE design with a compact active parameter footprint (roughly 4B active). Gemma achieves higher aggregate benchmark marks (76.21 on average), but its 26B total parameter footprint demands around 18 GB of VRAM even at 4-bit quantization, ruling out 16GB consumer GPUs entirely. Villanova2 fits within a far more accessible hardware envelope (under 11 GB at Q4_K_M), ships under a completely unrestrictive Apache 2.0 license, and provides complete training source documentation, making it significantly easier to audit for commercial deployment.
Energy cost on Intel Arc A770 16GB (~66 tok/s, Q4_K_M) vs flagship API pricing.
| Source | Cost per 1M tokens |
|---|---|
Local (energy only)Villanova2-15B-A3B-Base on Intel Arc A770 16GB · ~66 tok/s · 225W | $0.114 |
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 14 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA GeForce RTX 5060 TiVast.ai · Spot · 16 GB VRAM | $0.05 |
NVIDIA Tesla V100 16GBVast.ai · Spot · 16 GB VRAM | $0.08 |
NVIDIA RTX A4000Vast.ai · Spot · 16 GB VRAM | $0.09 |
NVIDIA Tesla V100 16GBVast.ai · On-Demand · 16 GB VRAM | $0.09 |
NVIDIA GeForce RTX 5060 TiVast.ai · On-Demand · 16 GB VRAM | $0.09 |
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.