M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models
FreeScalable test-time compute with Mamba reasoning models
FreeFree tier
Inputs: text
About M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models
M1 is a hybrid linear RNN reasoning model built on the Mamba architecture, designed for scalable test-time compute in mathematical reasoning. It leverages distillation from existing reasoning models and reinforcement learning to enhance performance. M1 achieves over 3x speedup compared to same-size transformer models on vLLM and matches the accuracy of DeepSeek R1 distilled reasoning models on AIME and MATH benchmarks. By using self-consistency voting, it improves accuracy under fixed generation time budgets, offering memory-efficient inference with linear computational complexity.
Key Features
Hybrid linear RNN architecture built on Mamba for memory-efficient inference
Distillation from existing reasoning models combined with RL training
Over 3x speedup vs same-size transformer models on vLLM
Matches performance of DeepSeek R1 distilled reasoning models at similar scale
Self-consistency voting for higher accuracy under fixed generation time budgets
Open-source code available
Pros & Cons
Pros
- Memory-efficient linear RNN scales better to long contexts than quadratic transformers
- Faster inference speed (3x vs same-size transformer) enables higher throughput
- Competitive accuracy with state-of-the-art transformer reasoning models
- Open-source model available for research and deployment
Cons
- Requires specialized Mamba architecture hardware/software support
- Not as extensively benchmarked as transformer-based reasoning models
- Distillation process may limit originality compared to full RL training from scratch
Best For
Mathematical problem solving (AIME, MATH benchmarks)Long chain-of-thought reasoningScalable test-time compute for complex reasoning tasks
FAQ
What is M1?
M1 is a hybrid linear RNN reasoning model built on the Mamba architecture, designed for scalable test-time compute in mathematical reasoning.
How does M1 compare to transformer models?
M1 achieves over 3x speedup compared to same-size transformer models on vLLM and matches the accuracy of DeepSeek R1 distilled reasoning models on AIME and MATH benchmarks.
What training methods are used for M1?
M1 leverages a distillation process from existing reasoning models and is further enhanced through reinforcement learning (RL) training.
Is the code for M1 available?
Yes, the code is available at the provided GitHub link in the arXiv paper.
What benchmarks were used to evaluate M1?
M1 was evaluated on the AIME and MATH benchmarks for mathematical reasoning.