diff --git a/packages/models/src/models/alibaba.ts b/packages/models/src/models/alibaba.ts
index e50d18610a..df69908bd8 100644
--- a/packages/models/src/models/alibaba.ts
+++ b/packages/models/src/models/alibaba.ts
@@ -1782,6 +1782,43 @@ export const alibabaModels = [
 					"reasoning_effort",
 				],
 			},
+			{
+				providerId: "novita",
+				externalId: "qwen/qwen3.8-max",
+				inputPrice: "2e-6",
+				cachedInputPrice: "0.25e-6",
+				outputPrice: "6e-6",
+				requestPrice: "0",
+				contextSize: 1000000,
+				maxOutput: 131072,
+				// novita accepts every reasoning_effort tier but none of them changes
+				// the deployment's behaviour (thinking stays on even for "none"), so
+				// no tier is declared and reasoning_effort is left out of
+				// supportedParameters below
+				reasoning: true,
+				reasoningOutput: "omit",
+				streaming: true,
+				vision: true,
+				tools: true,
+				// Qwen thinking models reject tool_choice "required" or object
+				supportedToolChoices: ["auto", "none"],
+				jsonOutput: true,
+				jsonOutputSchema: true,
+				// novita's deployment 400s on the developer role
+				supportsDeveloperRole: false,
+				supportedParameters: [
+					"temperature",
+					"max_tokens",
+					"top_p",
+					"frequency_penalty",
+					"presence_penalty",
+					"stop",
+					"stream",
+					"tools",
+					"tool_choice",
+					"response_format",
+				],
+			},
 			{
 				providerId: "scx-ai-gp",
 				externalId: "qwen3.8-max",
diff --git a/packages/models/src/models/moonshot.ts b/packages/models/src/models/moonshot.ts
index 6ca6896725..b755a80969 100644
--- a/packages/models/src/models/moonshot.ts
+++ b/packages/models/src/models/moonshot.ts
@@ -455,9 +455,9 @@ export const moonshotModels = [
 			{
 				providerId: "novita",
 				externalId: "moonshotai/kimi-k2.6",
-				inputPrice: "0.95e-6",
+				inputPrice: "0.8e-6",
 				cachedInputPrice: "0.16e-6",
-				outputPrice: "4.0e-6",
+				outputPrice: "3.4e-6",
 				requestPrice: "0",
 				contextSize: 262144,
 				maxOutput: 262144,
@@ -601,6 +601,29 @@ export const moonshotModels = [
 					"reasoning_effort",
 				],
 			},
+			{
+				providerId: "novita",
+				externalId: "moonshotai/kimi-k2.7-code",
+				inputPrice: "0.95e-6",
+				cachedInputPrice: "0.19e-6",
+				outputPrice: "4.0e-6",
+				requestPrice: "0",
+				contextSize: 262144,
+				maxOutput: 262144,
+				reasoning: true,
+				// Thinking is always on for kimi-k2.7-code: novita accepts `none` and
+				// `minimal` but keeps reasoning, so only low..max are offered.
+				reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
+				streaming: true,
+				vision: true,
+				tools: true,
+				// novita 400s on forced tool_choice for this deployment
+				supportedToolChoices: ["auto", "none"],
+				jsonOutput: true,
+				jsonOutputSchema: true,
+				// novita's deployment 400s on the developer role
+				supportsDeveloperRole: false,
+			},
 			{
 				providerId: "nebius",
 				externalId: "moonshotai/Kimi-K2.7-Code",
diff --git a/packages/models/src/models/tencent.ts b/packages/models/src/models/tencent.ts
index f24e53b921..09f5070c03 100644
--- a/packages/models/src/models/tencent.ts
+++ b/packages/models/src/models/tencent.ts
@@ -41,6 +41,8 @@ export const tencentModels = [
 				reasoningEfforts: ["none", "low", "high"],
 				vision: false,
 				tools: true,
+				// novita 400s on a named function choice for this deployment
+				supportedToolChoices: ["auto", "none", "required"],
 				jsonOutputSchema: true,
 			},
 		],
diff --git a/packages/models/src/models/zai.ts b/packages/models/src/models/zai.ts
index 6354994410..8f856b15b6 100644
--- a/packages/models/src/models/zai.ts
+++ b/packages/models/src/models/zai.ts
@@ -27,6 +27,33 @@ export const zaiModels = [
 				webSearchPrice: "0.01",
 				jsonOutput: true,
 			},
+			{
+				providerId: "novita",
+				externalId: "zai-org/glm-5.2",
+				inputPrice: "1.4e-6",
+				cachedInputPrice: "0.26e-6",
+				outputPrice: "4.4e-6",
+				requestPrice: "0",
+				contextSize: 1048576,
+				maxOutput: 131072,
+				streaming: true,
+				reasoning: true,
+				reasoningEfforts: [
+					"none",
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh",
+					"max",
+				],
+				vision: false,
+				tools: true,
+				// JSON mode is unreliable on this deployment: roughly half of the
+				// responses put the object in reasoning_content with empty content, and
+				// the rest markdown-fence it. json_schema behaves the same way.
+				jsonOutput: false,
+			},
 			{
 				providerId: "canopywave",
 				test: "skip", // over-reasons heavily and streams slowly (~1 tok/s), so the 60s streaming timeout is flaky
@@ -235,36 +262,33 @@ export const zaiModels = [
 			{
 				providerId: "novita",
 				externalId: "zai-org/glm-5.1",
-				inputPrice: "1.4e-6",
+				inputPrice: "1.38e-6",
 				cachedInputPrice: "0.26e-6",
 				outputPrice: "4.4e-6",
 				requestPrice: "0",
 				contextSize: 204800,
-				maxOutput: 131100,
+				maxOutput: 131072,
 				quantization: "fp8",
 				streaming: true,
 				reasoning: true,
+				reasoningEfforts: [
+					"none",
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh",
+					"max",
+				],
 				// novita's glm-5.1 reasons adaptively and omits reasoning_content
 				// for simple prompts; no parameter forces it on
 				reasoningOutput: "omit",
 				vision: false,
 				tools: true,
 				jsonOutput: true,
-				// novita disables thinking when reasoning_effort is forwarded
-				// (empty reasoning_content); omitting it reasons by default, so
-				// exclude reasoning_effort here (verified live 2026-07-14)
-				supportedParameters: [
-					"temperature",
-					"max_tokens",
-					"top_p",
-					"frequency_penalty",
-					"presence_penalty",
-					"stop",
-					"stream",
-					"response_format",
-					"tools",
-					"tool_choice",
-				],
+				// JSON mode consistently markdown-fences the object on this
+				// deployment, so normalize it defensively in both modes.
+				healStreamingJsonOutput: true,
 			},
 			{
 				providerId: "together-ai",
@@ -398,7 +422,7 @@ export const zaiModels = [
 				outputPrice: "3.2e-6",
 				requestPrice: "0",
 				contextSize: 202800,
-				maxOutput: 131100,
+				maxOutput: 131072,
 				quantization: "fp8",
 				streaming: true,
 				reasoning: true,