Sapient Intelligence Releases HRM-Text 1B

Sapient Intelligence has unveiled HRM-Text 1B, a model applying the Hierarchical Reasoning approach. This model is notable for showing excellent performance on certain reasoning tasks despite using very little training data and cost.

Key Training Metrics:

• Parameter count: 1B
• Training data: 40B tokens (about 1/100 to 1/900 the data used by existing models)
• Training cost: approximately $1,000
• Training duration: 1.9 days on 16 GPUs

Benchmark Performance: HRM-Text 1B showed strength on benchmarks requiring multi-step reasoning.

• MATH: 56.2 (surpassing Llama3.2 3B's 48.0)
• DROP: 82.2 (surpassing Llama3.2 3B's 45.2)

However, due to the limited amount of training data, it recorded lower performance than Qwen3.5 2B or Olmo3 7B on MMLU (60.7), which measures knowledge recall ability. This shows that 40B tokens was insufficient for the model to acquire vast world knowledge.

Note: These figures are self-reported by the developer, and given the nature of the MATH and DROP benchmarks, further verification may be needed regarding the possibility of data contamination.