diff --git a/packages/catalog/CHANGELOG.md b/packages/catalog/CHANGELOG.md
index feda2a158..30b6b1c0b 100644
--- a/packages/catalog/CHANGELOG.md
+++ b/packages/catalog/CHANGELOG.md
@@ -14,6 +14,7 @@
 - Fixed forced `tool_choice` 400s (`tool_choice 'specified' is incompatible with thinking enabled`) on Kimi Code's Anthropic-compatible endpoint for the `kimi-for-coding`, `kimi-for-coding-highspeed`, and `k3` aliases: the Anthropic-surface compat matcher only recognised Moonshot's native `kimi-k2.7-code*` ids, so thinking-locked kimi-code models kept `supportsForcedToolChoice: true` and the forced selector was sent to a host that always thinks. These models now resolve `requiresThinkingEnabled`, keeping thinking on and downgrading forced choices to `auto`.
 - Retried empty successful provider discovery responses after the short non-authoritative interval instead of caching them for the full catalog TTL ([#6620](https://github.com/can1357/oh-my-pi/issues/6620)).
 - Fixed GitHub Copilot Claude models with no bundled catalog reference (e.g. a freshly served `claude-opus-5`) discovering with `reasoning: false`/`thinking: null` and no effort dial, and disappearing along with their synthesized `-1m` sibling on offline reads: reference-less Copilot models on the anthropic-messages proxy now derive the adaptive reasoning ladder from the model id, and the cache restores their compile-time `COPILOT_API_HEADERS` by value instead of dropping them as unrestorable ([#6664](https://github.com/can1357/oh-my-pi/issues/6664)).
+- Fixed Kimi Code (`kimi-code`) reporting `maxTokens: 32000` for every model — its `/coding/v1/models` discovery mapper and the bundled catalog applied a blanket constant, truncating `k3`/`k3-256k` output at ~4x below their real 131072 ceiling and `kimi-for-coding`/`kimi-for-coding-highspeed` below their 32768 ceiling. Output caps are now derived per family, and the model cache is invalidated so upgrades drop the stale `maxTokens: 32000` rows (including the discovery-only `k3-256k`) instead of serving them until the next network refresh ([#6711](https://github.com/can1357/oh-my-pi/issues/6711)).
 
 ## [17.1.3] - 2026-07-24
 
diff --git a/packages/catalog/scripts/generate-models.ts b/packages/catalog/scripts/generate-models.ts
index 6971eab37..48cecef63 100644
--- a/packages/catalog/scripts/generate-models.ts
+++ b/packages/catalog/scripts/generate-models.ts
@@ -37,6 +37,7 @@ import {
 	clampKimiK27CodeMaxTokens,
 	isFireworksKimiK2ModelId,
 	isKimiK27CodeModelId,
+	kimiCodeMaxTokens,
 	META_MUSE_STATIC_MODELS,
 	MODELS_DEV_PROVIDER_DESCRIPTORS,
 	mapModelsDevToModels,
@@ -304,6 +305,12 @@ function applyKimiMaxTokensCap(models: readonly ModelSpec[]): ModelSpec[] {
 			const capped = clampKimiK27CodeMaxTokens(model.id, model.maxTokens);
 			return capped === model.maxTokens ? model : { ...model, maxTokens: capped };
 		}
+		if (model.provider === "kimi-code") {
+			// Discovery snapshots carried maxTokens=32000 uniformly (#6711); pin the
+			// documented per-family output ceilings and leave legacy K2 rows as-is.
+			const capped = kimiCodeMaxTokens(model.id, model.maxTokens);
+			return capped === model.maxTokens ? model : { ...model, maxTokens: capped };
+		}
 		return model;
 	});
 }
diff --git a/packages/catalog/src/model-cache.ts b/packages/catalog/src/model-cache.ts
index 871a57406..da10305f9 100644
--- a/packages/catalog/src/model-cache.ts
+++ b/packages/catalog/src/model-cache.ts
@@ -9,8 +9,10 @@ import type { Api, Model, ModelSpec } from "./types";
 // Rows persist ModelSpec JSON (sparse `compat`, never the resolved record);
 // the model manager rebuilds via `buildModel` on load. Request headers are
 // intentionally omitted: arbitrary provider-defined header names can carry
-// credentials. v11 invalidates rows that may persist derived computer-use
-// support without provenance; v10 deletes rows that may contain persisted
+// credentials. v12 invalidates Kimi Code rows carrying the blanket
+// maxTokens: 32000 that predate per-family output caps (k3/k3-256k -> 131072,
+// kimi-for-coding[-highspeed] -> 32768, #6711); v11 invalidates rows that may
+// persist derived computer-use
 // headers and records which model ids lost headers or cannot be rebuilt.
 // v9 invalidated Kimi Code rows predating live effort and protocol metadata;
 // v8 invalidated Codex discovery rows predating provider-native V2 compaction
@@ -20,7 +22,7 @@ import type { Api, Model, ModelSpec } from "./types";
 // retired unknown-limit sentinels (222222/8888); v5 invalidated rows predating
 // effort-tier variant collapsing (raw `-low`/`-high`/`-thinking` member ids);
 // v4 dropped the pre-efforts ThinkingConfig shape.
-const CACHE_SCHEMA_VERSION = 11;
+const CACHE_SCHEMA_VERSION = 12;
 const HEADER_RESTORE_VERSION = 1;
 
 interface CacheRow {
diff --git a/packages/catalog/src/models.json b/packages/catalog/src/models.json
index c3558a915..d89afe2f4 100644
--- a/packages/catalog/src/models.json
+++ b/packages/catalog/src/models.json
@@ -35139,7 +35139,7 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 1048576,
-			"maxTokens": 32000,
+			"maxTokens": 131072,
 			"thinking": {
 				"mode": "effort",
 				"efforts": [
@@ -35213,7 +35213,7 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 262144,
-			"maxTokens": 32000,
+			"maxTokens": 32768,
 			"thinking": {
 				"mode": "effort",
 				"efforts": [
@@ -35248,7 +35248,7 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 262144,
-			"maxTokens": 32000,
+			"maxTokens": 32768,
 			"thinking": {
 				"mode": "effort",
 				"efforts": [
diff --git a/packages/catalog/src/provider-models/openai-compat.ts b/packages/catalog/src/provider-models/openai-compat.ts
index 1b57bd811..ce8222e13 100644
--- a/packages/catalog/src/provider-models/openai-compat.ts
+++ b/packages/catalog/src/provider-models/openai-compat.ts
@@ -2707,6 +2707,34 @@ function mapKimiApiFormat(protocol: unknown): OpenAICompat["kimiApiFormat"] {
 	return undefined;
 }
 
+/**
+ * Kimi Code output ceilings by model family. The `/coding/v1/models` discovery
+ * envelope carries no output-limit field, so the mapper supplies the documented
+ * per-family caps instead of a blanket constant. Values match models.dev's
+ * `kimi-for-coding` and `moonshotai` kimi-k3 entries. See #6711.
+ */
+export const KIMI_CODE_K3_MAX_TOKENS = 131_072;
+export const KIMI_CODE_FOR_CODING_MAX_TOKENS = 32_768;
+
+/** Fallback output cap for Kimi Code families without a documented ceiling (legacy K2 discovery rows). */
+export const KIMI_CODE_DEFAULT_MAX_TOKENS = 32_000;
+
+/**
+ * Resolve a Kimi Code model's output ceiling from its id: `k3` / `k3-256k` ->
+ * 131072, `kimi-for-coding[-highspeed]` -> 32768, everything else -> `fallback`.
+ */
+export function kimiCodeMaxTokens(modelId: string, fallback?: number): number;
+export function kimiCodeMaxTokens(modelId: string, fallback: number | null): number | null;
+export function kimiCodeMaxTokens(
+	modelId: string,
+	fallback: number | null = KIMI_CODE_DEFAULT_MAX_TOKENS,
+): number | null {
+	const id = modelId.toLowerCase();
+	if (id.startsWith("k3")) return KIMI_CODE_K3_MAX_TOKENS;
+	if (id.startsWith("kimi-for-coding")) return KIMI_CODE_FOR_CODING_MAX_TOKENS;
+	return fallback;
+}
+
 export function kimiCodeModelManagerOptions(
 	config?: KimiCodeModelManagerConfig,
 ): ModelManagerOptions<"openai-completions"> {
@@ -2739,7 +2767,7 @@ export function kimiCodeModelManagerOptions(
 							reasoning,
 							input: entry.supports_image_in === true || id.includes("k2.5") ? ["text", "image"] : ["text"],
 							contextWindow: typeof entry.context_length === "number" ? entry.context_length : 262144,
-							maxTokens: 32000,
+							maxTokens: kimiCodeMaxTokens(id),
 							thinking,
 							compat: {
 								thinkingFormat: thinking ? "kimi" : "zai",
diff --git a/packages/catalog/test/kimi-code-provider.test.ts b/packages/catalog/test/kimi-code-provider.test.ts
index 50ea839e8..3a3534cb2 100644
--- a/packages/catalog/test/kimi-code-provider.test.ts
+++ b/packages/catalog/test/kimi-code-provider.test.ts
@@ -35,6 +35,7 @@ describe("Kimi Code provider catalog", () => {
 			name: "K3",
 			reasoning: true,
 			contextWindow: 1_048_576,
+			maxTokens: 131_072,
 			thinking: {
 				mode: "effort",
 				efforts: [Effort.Low, Effort.High, Effort.Max],
@@ -67,6 +68,28 @@ describe("Kimi Code provider catalog", () => {
 		expect(legacy?.thinking?.efforts).toEqual([Effort.Minimal, Effort.Low, Effort.Medium, Effort.High]);
 	});
 
+	it("derives per-family output caps instead of a blanket constant (#6711)", async () => {
+		const models = await discover([
+			LIVE_K3,
+			{ ...LIVE_K3, id: "k3-256k", display_name: "K3 256k", context_length: 262_144 },
+			{ id: "kimi-for-coding", display_name: "K2.7 Code", context_length: 262_144, supports_reasoning: true },
+			{
+				id: "kimi-for-coding-highspeed",
+				display_name: "K2.7 Code Highspeed",
+				context_length: 262_144,
+				supports_reasoning: true,
+			},
+			{ id: "kimi-k2", display_name: "Kimi K2", context_length: 262_144 },
+		]);
+		const maxTokensFor = (id: string) => models.find(model => model.id === id)?.maxTokens;
+
+		expect(maxTokensFor("k3")).toBe(131_072);
+		expect(maxTokensFor("k3-256k")).toBe(131_072);
+		expect(maxTokensFor("kimi-for-coding")).toBe(32_768);
+		expect(maxTokensFor("kimi-for-coding-highspeed")).toBe(32_768);
+		expect(maxTokensFor("kimi-k2")).toBe(32_000);
+	});
+
 	it("lets supports_thinking_type override the legacy reasoning flag", async () => {
 		const models = await discover([
 			{ ...LIVE_K3, id: "non-thinking", supports_thinking_type: "no", think_efforts: undefined },