feat: upgrade MiniMax default model to M3

[octo-patch]

Summary

Upgrade the MiniMax LLM provider integration to use the latest MiniMax-M3 as the default model. MiniMax-M3 features a 512K context window, up to 128K output, and image input support.

Changes

• Add MiniMax-M3 to the model selection list and set as the default for MiniMaxAPI
• Keep MiniMax-M2.7 and add MiniMax-M2.7-highspeed as alternatives
• Remove older models (MiniMax-M2.5, MiniMax-M2.5-highspeed)
• Update related unit tests and integration tests

Model Lineup (after this PR)

Model	Context	Notes
MiniMax-M3	512K	Default, supports image input
MiniMax-M2.7	204K	Previous generation
MiniMax-M2.7-highspeed	204K	Previous generation, low-latency variant

File Changes

File	Change
opencompass/models/minimax_api.py	MiniMaxAPI default path → MiniMax-M3, default max_seq_len → 524288
opencompass/configs/models/minimax/minimax_m3.py	New — M3 config (default)
opencompass/configs/models/minimax/minimax_m2_7_highspeed.py	New — M2.7-highspeed config
opencompass/configs/models/minimax/minimax_m2_5.py	Removed
opencompass/configs/models/minimax/minimax_m2_5_highspeed.py	Removed
tests/models/test_minimax_api.py	Update tests for new default and lineup

Usage

from opencompass.models import MiniMaxAPI

# Uses MiniMax-M3 by default (512K context)
model = MiniMaxAPI(key="ENV")  # reads from MINIMAX_API_KEY

# Or specify explicitly
model_m27 = MiniMaxAPI(path="MiniMax-M2.7", key="ENV")

results = model.generate(["What is 2+2?"], max_out_len=500)

Test plan

• Unit tests updated for M3 as default model
• Integration tests updated (test_real_api_call uses M3, test_real_api_m2_7_highspeed validates highspeed variant)
• Backward compatibility tests retained for MiniMaxChatCompletionV2