v0.6.71 — minimax-m2.7 alias + dogfood launcher

What's new

• minimax-m2.7 alias → mlx-community/MiniMax-M2.7-4bit-mxfp4 (parser/reasoning = minimax, dense, spec-decode supported). MiniMax-M2.7 is the same family as M2.5 but with active 10B params for faster decoding.
• scripts/run_dogfood_mvp.sh — single-Mac dogfood launcher. Brings up rapid-mlx serve behind a Cloudflare tunnel (quick or named) for hands-on use behind a public router. Defaults to --no-thinking --cors-origins * so browser-based chat UIs (Open WebUI, LobeChat, Big-AGI) work without CORS friction or a <think> prelude.

Try it

pip install --upgrade rapid-mlx
rapid-mlx info minimax-m2.7         # confirm alias resolves
rapid-mlx serve minimax-m2.7        # serve via OpenAI-compatible API

For the dogfood tunnel:

scripts/run_dogfood_mvp.sh start    # auto-detects ~/.cloudflared/config.yml

Release validation

Gate	Result
make release-smoke (clean venv wheel install + import)	PASS
make smoke (lint + audit + 4039 unit tests)	PASS
make stress (8 scenarios incl. tool-call storm)	8/8 PASS
make check (doctor gated tier vs qwen3.5-35b)	6/6 PASS
Server latency 10-seq @ qwen3.5-4b	109.5 tok/s, CV 7.4%
Integrations: Anthropic SDK	5/5 PASS
Integrations: smolagents	4/4 PASS
Integrations: pydantic_ai	5/6 PASS (1 expected small-model max_tokens overflow)
pr_validate (3 review rounds converged on real fixes; round 4 hit DeepSeek's documented asymptotic-NIT spiral)	accepted

🤖 Generated with Claude Code