Monolith-1.0 · Released July 17, 2026
Monolith.
The frontier is open. Monolith-1.0 leads every major reasoning benchmark, and its weights are on Hugging Face.
First place,
on everything that matters.
Reasoning, math, science, and long-context analysis. Evaluation protocols and reproduction details are published alongside every release.
Benchmark
Monolith-1
GPT-5.4
Claude Opus 4.6
Gemini 3.1 Pro
Kimi K2.6
Humanity's Last Exam
99.4%
40.0%
40.0%
44.4%
34.7%
AIME 2025
100%
99.2%
96.7%
98.3%
96.4%
GPQA Diamond
95.9%
92.8%
91.3%
94.3%
90.5%
MMLU-Pro
96.2%*
90.2%
89.7%
90.8%
88.9%
* Evaluated using Basalt's internal harness. Full methodology in the technical report.
Model card
Specifications.
Full model card, evaluations, and training details available in the paper.
Parameters
1.6T total / 49B active
Architecture
Mixture of Experts, 128 experts
Context window
1,048,576 tokens
Training compute
1.8 × 10²⁵ FLOPs
Training data cutoff
December 2025
License
MIT
Availability