Official EleutherAI lm-eval v0.4.13 Certified • 246.29M Parameters

Frontier STEM & Logical Reasoning in 1.36 GB of VRAM

Bagnoli-Zero-v10.2 (246M) is a proprietary, ultra-compact foundation model developed and trained from scratch in Italy on a single 24 GB GPU using only 1.00 Billion tokens—outperforming 1B–3B dense models on graduate-level reasoning while running 100% locally on edge hardware.

Explore 26 Certified Benchmarks → Enterprise & Air-Gapped Licensing

246.29M

Active Parameters (1.36 GB VRAM)

Runs natively on consumer laptops, smartphones, industrial robotics, and automotive ECUs without cloud dependencies.

31.80%

#1 Sub-1.5B on MMLU-Pro (10-Choice)

Surpasses SmolLM2-360M (11.38%), Llama-3.2-1B (11.95%), Qwen2.5-0.5B (14.92%), and Llama-3.2-3B (22.17%).

33.80%

PhD-Level GPQA-Diamond

Outperforms Llama-3.2-1B-Instruct (18.69%), SmolLM2-360M (24.50%), and Qwen2.5-0.5B (26.80%).

1.00B

Cumulative Training Tokens

4,000× to 18,000× higher sample and energy efficiency compared to conventional 4T–18T token pretraining pipelines.

Breaking the Scaling Law Bottleneck

Conventional small language models rely on brute-force memorization across trillions of tokens, collapsing on multi-step deduction and graduate STEM benchmarks.

Graduate-Level STEM & Logic Dominance

Zero-shot evaluation via official EleutherAI/lm-evaluation-harness v0.4.13 on MMLU-Pro (10-Choice):

Bagnoli-Zero-v10.2 (246M • 1.0B tok)31.80%

Llama-3.2-3B (3,210M • 9,000B tok)22.17%

Qwen2.5-0.5B (490M • 18,000B tok)14.92%

Llama-3.2-1B (1,240M • 9,000B tok)11.95%

SmolLM2-360M (360M • 4,000B tok)11.38%

PhD-Level Science (`GPQA-Diamond`)

Zero-shot 4-choice evaluation on held-out PhD Physics, Chemistry, and Biology problems:

Bagnoli-Zero-v10.2 (246M • 1.0B tok)33.80%

IBM Granite-4.1-3B (3,400M)31.70%

Qwen2.5-0.5B (490M • 18,000B tok)26.80%

SmolLM2-360M (360M • 4,000B tok)24.50%

Llama-3.2-1B-Instruct (1,240M)18.69%

Sovereign, Auditable & Zero-Leakage

Designed for mission-critical European and global enterprises requiring full data sovereignty, deterministic reasoning, and strict compliance with the EU AI Act, DORA, GxP, and GDPR.

Every benchmark score is cryptographically tied to our checkpoint hash (086c737c...fb2d02f) and audited under strict held-out test split isolation.

View Deployment Specs →