From d2e7d8b7c1d31d3661f769d5caef7ba8d50c3e03 Mon Sep 17 00:00:00 2001
From: David Brailovsky <davidbrai@gmail.com>
Date: Mon, 19 Oct 2026 11:35:34 +0200
Subject: [PATCH] feat(ai): add MiniMax-M3.1-Flash and hide only old MiniMax
 models

Replace the direct minimax allowlist with a denylist of the older M2, M2.1,
M2.5, and M2.5-highspeed models, applied where MiniMax models are built.
MiniMax models with models.dev effort levels use adaptive thinking with
output_config.effort and mark thinking off as unsupported, since
MiniMax-M3.1-Flash rejects thinking.type=disabled.

closes #11121
---
 packages/ai/CHANGELOG.md                      |  1 +
 packages/ai/scripts/generate-models.ts        | 21 +++++++------------
 ...anthropic-adaptive-thinking-models.test.ts |  5 +++++
 3 files changed, 13 insertions(+), 14 deletions(-)

diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index d265f5280..c83a7c931 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -10,6 +10,7 @@
 
 - Added `AI_TELEMETRY_SCHEMA` and opt-in `pi.ai.request`/`pi.ai.poll` tracing around Models requests, with content-free first-event, block-size, and final usage attributes, plus duration, token, and USD cost metrics through `telemetry.metrics`.
 - Failed model request spans report `provider_error` status with up to 500 error-message characters; cancelled requests keep `ok` status with `stop: "aborted"`.
+- Added `MiniMax-M3.1-Flash` to the `minimax` and `minimax-cn` providers, with thinking levels low through max sent as `output_config.effort`. Thinking cannot be turned off for this model. New MiniMax models from models.dev now appear without a manual allowlist entry; only the older M2, M2.1, M2.5, and M2.5-highspeed models stay hidden ([#11121](https://github.com/earendil-works/pi/issues/11121))
 
 ### Fixed
 
diff --git a/packages/ai/scripts/generate-models.ts b/packages/ai/scripts/generate-models.ts
index 29480825e..384baa913 100644
--- a/packages/ai/scripts/generate-models.ts
+++ b/packages/ai/scripts/generate-models.ts
@@ -2405,12 +2405,17 @@ async function loadModelsDevData(): Promise<Model<any>[]> {
 			{ key: "minimax", provider: "minimax", baseUrl: "https://api.minimax.io/anthropic" },
 			{ key: "minimax-cn", provider: "minimax-cn", baseUrl: "https://api.minimaxi.com/anthropic" },
 		] as const;
+		// MiniMax asked to hide older models from its direct providers (#2596).
+		const minimaxHiddenIds = new Set(["MiniMax-M2", "MiniMax-M2.1", "MiniMax-M2.5", "MiniMax-M2.5-highspeed"]);
 
 		for (const { key, provider, baseUrl } of minimaxVariants) {
 			if (data[key]?.models) {
 				for (const [modelId, model] of Object.entries(data[key].models)) {
 					const m = model as ModelsDevModel;
-					if (m.tool_call !== true) continue;
+					if (m.tool_call !== true || minimaxHiddenIds.has(modelId)) continue;
+					// Models with effort levels (e.g. MiniMax-M3.1-Flash) take adaptive thinking with
+					// output_config.effort and reject thinking.type="disabled" (no "none" effort, so off is null).
+					const thinkingLevelMap = getEffortThinkingLevelMap(m.reasoning_options ?? []);
 
 					models.push({
 						id: modelId,
@@ -2424,8 +2429,8 @@ async function loadModelsDevData(): Promise<Model<any>[]> {
 						cost: getModelsDevCost(m.cost),
 						contextWindow: m.limit?.context || 4096,
 						maxTokens: m.limit?.output || 4096,
+						...(thinkingLevelMap ? { thinkingLevelMap, compat: { forceAdaptiveThinking: true } } : {}),
 					});
-					recordModelsDevReasoningOptions(provider, modelId, m);
 				}
 			}
 		}
@@ -3137,18 +3142,6 @@ async function generateModels() {
 		}
 	}
 
-	const minimaxDirectSupportedIds = new Set(["MiniMax-M2.7", "MiniMax-M2.7-highspeed", "MiniMax-M3"]);
-
-	for (let i = allModels.length - 1; i >= 0; i--) {
-		const candidate = allModels[i];
-		if (
-			(candidate.provider === "minimax" || candidate.provider === "minimax-cn") &&
-			!minimaxDirectSupportedIds.has(candidate.id)
-		) {
-			allModels.splice(i, 1);
-		}
-	}
-
 	// OpenAI Codex (ChatGPT OAuth) models
 	// NOTE: These are not fetched from models.dev; we keep a small, explicit list to avoid aliases.
 	// Older model limits are based on observed server behavior; GPT-5.6 and GPT-6 use Codex's 272k default catalog limit.
diff --git a/packages/ai/test/anthropic-adaptive-thinking-models.test.ts b/packages/ai/test/anthropic-adaptive-thinking-models.test.ts
index 69615cb07..a3a1c4414 100644
--- a/packages/ai/test/anthropic-adaptive-thinking-models.test.ts
+++ b/packages/ai/test/anthropic-adaptive-thinking-models.test.ts
@@ -12,6 +12,8 @@ const EXPECTED_CURRENT_ADAPTIVE_THINKING_MODELS = [
 	"fireworks/accounts/fireworks/models/deepseek-v4p1-flash",
 	"fireworks/accounts/fireworks/models/gpt-oss-120b",
 	"fireworks/accounts/fireworks/models/qwen3p8-max",
+	"minimax-cn/MiniMax-M3.1-Flash",
+	"minimax/MiniMax-M3.1-Flash",
 	"opencode/claude-opus-4-8",
 	"opencode/claude-opus-5",
 	"vercel-ai-gateway/anthropic/claude-opus-4.8",
@@ -38,6 +40,9 @@ describe("Anthropic adaptive thinking model metadata", () => {
 					// Regression for #9323: Fireworks uses catalog effort metadata and
 					// verified fallbacks, not a fixed set of adaptive model names.
 					modelId.startsWith("fireworks/") ||
+					// #11121: MiniMax models with catalog effort levels require adaptive thinking.
+					modelId.startsWith("minimax/") ||
+					modelId.startsWith("minimax-cn/") ||
 					/(opus[-.](4[-.][678]|5)|sonnet[-.]4[-.]6|sonnet[-.]5|haiku[-.]5[-.]5|fable[-.]5)/.test(modelId),
 			),
 		);