feat: upgrade MiniMax default model to M3 (512K context, 128K output)

[octo-patch]

Summary

This PR adds first-class support for the MiniMax-M3 architecture in OpenMythos:

• minimax_m3_config() — a new MythosConfig factory in open_mythos/variants.py whose structural dimensions mirror MiniMax-M3's design: dim=6144, 48 query / 8 KV heads (6× GQA), head_dim=128 via MLA rope+nope split (64+64), 32 routed + 2 shared experts with top-4 activation, 524 288-token (512K) context, 131 072-token (128K) max output, and rope_theta=10_000_000 for long-context stability.
• MINIMAX_M3_MODEL_ID — a constant ("MiniMaxAI/MiniMax-M3") added to open_mythos/tokenizer.py for convenient use with MythosTokenizer, whose vocab_size=200064 matches the config.
• Both symbols are re-exported from the top-level open_mythos package.
• tests/test_minimax_m3.py — unit tests covering config dimensions (including the new 512K context window and 128K max output), MoE structure (32 routed + 2 shared + top-4), MLA cache compression, LTI spectral radius stability, high rope_theta numerical safety, and full forward-pass / generation correctness with synthetic tensors on CPU.

All new tests are designed to run on CPU with synthetic tensors. No changes to existing code paths.

Why MiniMax-M3

MiniMax-M3 is the latest MiniMax release, with a 512K context window, up to 128K output tokens, and image-input support. Replacing the earlier draft (M2.7) with M3 keeps OpenMythos's reference config aligned with the current default model.

Usage

from open_mythos import minimax_m3_config, OpenMythos, MINIMAX_M3_MODEL_ID
import torch

cfg = minimax_m3_config()
model = OpenMythos(cfg)
ids = torch.randint(0, cfg.vocab_size, (1, 16))
logits = model(ids, n_loops=4)  # (1, 16, 200064)

# To use the MiniMax-M3 tokenizer:
from open_mythos.tokenizer import MythosTokenizer
tok = MythosTokenizer(MINIMAX_M3_MODEL_ID)

[Qodo-Free-For-OSS]

Hi, open_mythos.init.all includes "load_tokenizer" and "get_vocab_size", but neither symbol is imported into the package nor defined anywhere in the open_mythos package, causing AttributeError (and potentially failing from open_mythos import *).

Severity: action required | Category: correctness

How to fix: Define or remove exported names

Agent prompt to fix - you can give this to your LLM of choice:

Issue description

open_mythos/__init__.py lists load_tokenizer and get_vocab_size in __all__, but these symbols are not available on the package module. This breaks star-imports and any code expecting these public APIs.

Issue Context

The package currently imports MythosTokenizer and constants from open_mythos.tokenizer, but does not provide load_tokenizer/get_vocab_size wrappers.

Fix Focus Areas

• open_mythos/init.py[18-58]
• open_mythos/tokenizer.py[1-80]

Suggested implementation direction

Either:

1. Remove load_tokenizer and get_vocab_size from __all__ if they are not supported.

Or:
2) Implement load_tokenizer(model_id: str = DEFAULT_MODEL_ID) -> MythosTokenizer and get_vocab_size(model_id: str = DEFAULT_MODEL_ID) -> int in open_mythos/tokenizer.py, then import them in open_mythos/__init__.py and keep them in __all__.

---

Found by Qodo code review. FYI, Qodo is free for open-source.