PolicyEngine

MiniMax M3

MiniMax · elicited July 2026 independent labs · average interval-width rank #17.6 of 29 across the canonical panel · — (cost untracked).

minimax-m3

Canonical panel profile

One row per canonical quantity, each on its own scale: this model's pooled center and 90 percent interval, with the 29-model panel median as a gray tick.

Generation harness

From the paper's per-model disclosure (Appendix A16). The prompt and repeated-run design are identical across models; the output mechanism follows each provider's API surface.

Provider path
LiteLLM via OpenRouter
Output mechanism
forced JSON object (schema validated locally)
Completion budget
8000
Sampling
temperature 1.0
Reasoning
provider default
API identifier
openrouter/minimax/minimax-m3 (alias)

195 successful runs · elicited April and July 2026 · v4 prompts · 15 runs per model-quantity cell. Code · Raw responses · Paper (PDF)

Code and dataElicited April and July 2026 · 29 models · v4 prompts