From 1d06a1ca69335cb4c925b4f1092e1deb6d345bfb Mon Sep 17 00:00:00 2001 From: yuneng-jiang Date: Mon, 9 Nov 2026 11:03:39 -0800 Subject: [PATCH] test(e2e): repin the Together reasoning replay backend to MiniMax-M3.1-Flash (#50601) MiniMax-M3 reached its 2026-11-06 deprecation date in the cost map, which names MiniMax-M3.1-Flash as its successor --- tests/e2e/llm_translation/test_together_ai_e2e.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/tests/e2e/llm_translation/test_together_ai_e2e.py b/tests/e2e/llm_translation/test_together_ai_e2e.py index 2d5d5f2d93a..7775c5c6ed2 100644 --- a/tests/e2e/llm_translation/test_together_ai_e2e.py +++ b/tests/e2e/llm_translation/test_together_ai_e2e.py @@ -7,9 +7,9 @@ backends are pinned because the registry has no flag for what they prove: ``enab and the ``{"reasoning": {"enabled": false}}`` toggle that ``reasoning_effort="none"`` maps to are Qwen hybrid-model contracts, and the structured-output case runs on that hybrid model with reasoning off, since a reasoning-only model can spend the whole token budget -thinking and return no content. MiniMax-M3 is the serverless model whose +thinking and return no content. MiniMax-M3.1-Flash is the serverless model whose template renders a replayed ``reasoning_content`` back into the prompt (Qwen and -DeepSeek silently drop it). MiniMax-M3 honors that replayed field on nearly every call, not +DeepSeek silently drop it). MiniMax-M3.1-Flash honors that replayed field on nearly every call, not every call (one miss in dozens of otherwise identical calls), so the replay case asks up to ``REPLAY_ATTEMPTS`` times and fails only when no answer carries the secret, which a proxy that strips the field guarantees. Requires TOGETHER_API_KEY on the proxy; no @@ -56,7 +56,7 @@ from pydantic import BaseModel pytestmark = pytest.mark.e2e HYBRID_REASONING_BACKEND = "together_ai/Qwen/Qwen3.5-9B" -REASONING_REPLAY_BACKEND = "together_ai/MiniMaxAI/MiniMax-M3" +REASONING_REPLAY_BACKEND = "together_ai/MiniMaxAI/MiniMax-M3.1-Flash" SECRET_PROMPT = "Remember this for later and reply with just OK." SECRET_REASONING = "The user told me their favorite color is chartreuse. I must remember it."