MiniMax-M3 — Production Wiring (empirical)

The M3 companion to ../minimax/SKILL.md (M2.7). Every claim here was live-probed 2026-06-01 (fast subset re-verified 2026-06-23) on the Plus-High-Speed key. Full evidence + copy-paste snippets: ../../references/M3-EMPIRICAL.md.

Self-Evolving Skill: improves through use. If a flag stopped working, a limit moved, or the docs caught up with reality — fix this file + references/M3-EMPIRICAL.md immediately, don't defer. Re-verify with the scripts below before changing a documented fact.

The one rule: default to `reasoning_split: true`

M3 still emits <think>…</think> inside content by default (same footgun as M2.7). Setting reasoning_split: true moves the reasoning into a separate reasoning_content / reasoning_details field and leaves content clean — no regex stripping. This is the chosen default profile for everything migrating off M2.7-highspeed.

body = {
    "model": "MiniMax-M3",
    "messages": messages,
    "max_tokens": 4096,          # >= 1024 — thinking consumes budget before visible content
    "temperature": 0.2,
    "reasoning_split": True,     # clean content; reasoning in reasoning_content/_details
}
answer = resp["choices"][0]["message"]["content"]   # already clean — display directly

Need M2.7-highspeed-class speed on short/simple tasks? Add "reasoning": "disabled" (≈2× fewer tokens, ≈2× faster) — and keep the M2.7 <think> strip as a safety net, since "disabled" shortens but doesn't always remove the block. Keep thinking ON (default / "adaptive") for hard reasoning, coding, and agentic loops.

When to use M3 vs M2.7

Workload	Verdict
Clean chat / judgment / theory / JSON	✅ M3 + `reasoning_split:true` (the new default)
Short tagging / classification, latency-sensitive	✅ M3 + `reasoning:"disabled"`, or stay on plain `MiniMax-M2.7`
Vision (OCR, charts, screenshots)	✅ M3 only — M2.7 is text-only; M3 reads images correctly
Structured JSON	✅ M3 (`response_format` accepted) + `reasoning_split` + defensive parse
Long context (input up to ~1M)	✅ input accepts to ~1M, but reliable retrieval ≤ ~256K (400K now misses); 1M prefill ~235 s
Raw math / QP / risk on realistic N	❌ still route to Python (the M2.7 saturation guidance carries over)
Final deployable code	⚠️ scaffold-only; sandbox-validate (unchanged from M2.7)

Hard limits & gotchas (live-verified)

Input context ≈ 1,000,000 tokens (re-verified 2026-06-23). Accepts to ~1,000,180; 1,048,576 → context window exceeds limit. The docs' 1M now holds (was ~512K on 2026-06-01) — but 1M prefill is ~235 s and reliable needle retrieval is ≤ ~256K (400K misses 2/2). Operate at ≤ ~256K for retrieval-critical work.
Output max_tokens ≤ 524,288. > 524288 → invalid params … does not support max tokens > 524288 (raised from 512,000 — re-verified 2026-06-23).
n > 1 silently dropped. Was a hard 2013 rejection; now accepted-but-ignored — the response still carries exactly one choice. No true multi-sampling (re-verified 2026-06-23).
response_format accepted but not a hard JSON guarantee — M3 may still wrap with <think> / ```json fences / a trailing note. Pair with reasoning_split + try/except json.loads.
tool_choice forced did NOT compel a call in the trivial-prompt probe — re-test with a tool-relevant prompt before relying on forced tool calls.
No MiniMax-M3-highspeed on this key (2013 unknown model) despite docs.
All the M2.7 defensive snippets (<think> strip, base_resp rate-limit retry, cached-token reader) in ../minimax/SKILL.md apply unchanged.

Re-verify / detect drift

Scripts live at the plugin source checkout (scripts/ is stripped from the runtime cache), run from ~/eon/cc-skills/plugins/minimax:

export MINIMAX_API_KEY=...        # or rely on the 1Password op-path default (see `bun scripts/m3-cli.ts verify --help`)

bun scripts/m3-cli.ts verify          # fast drift check vs locked snapshot (exit 0/1/2)
bun scripts/m3-cli.ts probe [--out f] # full option/capability map (writes JSON; default m3_probe_results.json)
bun scripts/m3-cli.ts context-probe   # input-context ceiling + needle retrieval
bun scripts/m3-cli.ts bench           # speed/quality: default thinking vs reasoning:"disabled"
./scripts/minimax-check-upgrade       # catalog drift (lock includes MiniMax-M3)

Locked invariants: ../../references/fixtures/m3-capabilities-locked-2026-06-23.json. Schedule m3-cli.ts verify + minimax-check-upgrade (launchd template in templates/) to catch the day MiniMax ships M3-highspeed, opens up 1M context, or changes a limit.

Post-Execution Reflection

Locate yourself. — Confirm this is the canonical skills/m3/SKILL.md before editing.
What failed? — A flag that worked now errors, or vice-versa → fix here + M3-EMPIRICAL.md.
What drifted? — m3-cli verify flagged an invariant change → review, then bump the locked snapshot.
Log it. — Append to the Evolution log in references/M3-EMPIRICAL.md with trigger + fix + evidence.

m3