www.lesswrong.com/posts/QKDoZe6EKhxnFjLWK/is-mythos-good-at-cyber-because-it-kep...
1 correction found
a 230B-A20B Chinese model
Public MiniMax materials describe the M2 series as roughly 230B total parameters with about 10B activated parameters, not 20B.
Full reasoning
The publicly documented MiniMax M2 series is described as a ~230B-parameter MoE family with ~10B activated parameters, not 20B.
Evidence:
- The official MiniMax M2.1 post says the model is built on the previous M2 generation and "adopts a Mixture-of-Experts (MoE) architecture, with approximately 230B total parameters and ~10B activated parameters."
- The official MiniMax-M2 series technical report says: "The flagship M2 contains 229.9B total parameters with only 9.8B activated per token."
- The official MiniMax-M2.5 checkpoint is explicitly part of that same M2 series (its model card says MiniMax released M2, M2.1, and M2.5 over three and a half months), and the public checkpoint metadata shows about 228.7B total parameters.
So while the total-parameter part is approximately right, the A20B label appears to double the activated-parameter count. The public MiniMax documentation points to something like 230B-A10B, not 230B-A20B.
3 sources
- MiniMax M2.1: Post-Training Experience and Insights for Agent Models - MiniMax News | MiniMax
M2.1 is the latest flagship open-source model ... built upon further post-training optimization over the previous M2 generation. It adopts a Mixture-of-Experts (MoE) architecture, with approximately 230B total parameters and ~10B activated parameters.
- The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence
We introduce the MiniMax-M2 series ... The flagship M2 contains 229.9B total parameters with only 9.8B activated per token.
- MiniMaxAI/MiniMax-M2.5 · Hugging Face
The public checkpoint metadata for MiniMax-M2.5 shows safetensors total 228703644928 (about 228.7B parameters), and the model card describes M2.5 as part of the M2 series.