LLMs: 0-day fallbacks: mark type-blind vendors, never invent context

Author: enricoros

src/modules/llms/server/gemini/gemini.models.ts

@@ -1158,6 +1158,10 @@ export function geminiModelToModelDescription(geminiModel: GeminiWire_API_Models
     ? `🔗 ${knownModel?.labelOverride || displayName} -> ${knownModel.symLink}`
     : knownModel?.labelOverride || displayName;
 
+  // no '[?]' marker (evaluated 2026-08-14): API-characterized (supportedGenerationMethods + real ctx) - see llmsLabelUncurated
+  // if (!knownModel)
+  //   label = llmsLabelUncurated(label);
+
   // FIX: the Gemini 1114 model now returns 1121 as the version.. highlight the issue
   // if (geminiModel.name.endsWith('1114') && label.endsWith('1121'))
   //   label += ' (really: 1114)';

src/modules/llms/server/models.mappings.ts

@@ -9,12 +9,18 @@ import type { ModelDescriptionSchema } from './llm.server.types';
 // -- Uncurated-model label marker --
 
 /**
- * '[?] ' label prefix: marks a model no editorial table covers yet (0-day arrivals, unknown
- * variants). Writers: fromManualMapping 'super' resolution (below), nvidianim/modular 0-day
- * fallbacks. Readers: llm-registry-sync holds prefix-marked models with null contextWindow off
- * the publication push; listModels.test curated-context assertion; the website strips the
- * brackets and sinks the pub date (own regex, website repo - keep in sync on change).
- * Legacy variants not yet unified: bedrock ' [?]' label suffix, openai fallback bare '?'.
+ * '[?] ' label prefix - "the list API does not establish what this model IS".
+ * Not an "uncatalogued" badge: API-characterized 0-day arrivals stay unmarked.
+ *
+ * Mark:    type-blind catalogs (ids only), where a video/TTS/embedding id could masquerade as
+ *          chat - nvidianim, modular, sakanaai, moonshot, groq, deepseek, alibaba - plus the
+ *          'super' resolution below (unknown variant of a known family).
+ * Don't:   a type/modality filter proves chat - gemini, xai, together, novita, chutesai, cerebras.
+ *
+ * Readers: llm-registry-sync (marked + null contextWindow = held off the publication push),
+ *          listModels.test, the website (strips brackets, sinks pubDate; own regex, keep in sync).
+ *
+ * Legacy, not unified: bedrock ' [?]' suffix, openai bare '?'.
  */
 const LLM_LABEL_UNCURATED = '[?]';
 

src/modules/llms/server/openai/models/alibaba.models.ts

@@ -2,7 +2,7 @@ import { LLM_IF_OAI_Chat, LLM_IF_OAI_Fn, LLM_IF_OAI_Reasoning, LLM_IF_OAI_Vision
 
 import type { ModelDescriptionSchema } from '../../llm.server.types';
 
-import { fromManualMapping, llmsDefineManualMappings } from '../../models.mappings';
+import { fromManualMapping, llmsDefineManualMappings, llmsLabelUncurated } from '../../models.mappings';
 
 // --- Alibaba Model ID inference (auto-derived from _knownAlibabaChatModels) ---
 export type LlmsAlibabaModelId = typeof _knownAlibabaChatModels[number]['idPrefix'];
@@ -339,10 +339,9 @@ export function alibabaModelFilter(modelId: string): boolean {
 export function alibabaModelToModelDescription(alibabaModelId: string, created?: number): ModelDescriptionSchema {
   const md = fromManualMapping(_knownAlibabaChatModels, alibabaModelId, created, undefined, {
     idPrefix: alibabaModelId,
-    label: _alibabaFormatNewLabel(alibabaModelId),
+    label: llmsLabelUncurated(_alibabaFormatNewLabel(alibabaModelId)),
     description: 'Alibaba model (not yet curated).',
-    contextWindow: 131072,
-    maxCompletionTokens: 8192,
+    contextWindow: null, // the list API returns ids only - no context to trust
     interfaces: [LLM_IF_OAI_Chat, LLM_IF_OAI_Fn],
     hidden: true, // editorial policy: hide uncatalogued models by default (title-cased label only if ever shown)
   });

src/modules/llms/server/openai/models/cerebras.models.ts

@@ -119,14 +119,17 @@ function _cerebrasApiModelToFallback(model: WireCerebrasModel): KnownModel {
   const chatPrice = isFree ? { input: 'free' as const, output: 'free' as const } : { input: inputPerM, output: outputPerM };
 
   const isPreview = !!model.preview;
+  // catalog context or null, never a guess
+  // no '[?]' marker (evaluated 2026-08-14): API-characterized (modality filter + catalog capabilities) - see llmsLabelUncurated
+  const contextWindow = limits.max_context_length || null;
   const label = (model.name || model.id.replaceAll(/[_-]/g, ' ')) + (isPreview ? ' (Preview)' : '');
 
   return {
     idPrefix: model.id,
     isPreview,
     label,
     description: model.description || 'New Cerebras model.',
-    contextWindow: limits.max_context_length || 131072,
+    contextWindow,
     ...(limits.max_completion_tokens ? { maxCompletionTokens: limits.max_completion_tokens } : {}),
     interfaces,
     ...(parameterSpecs.length ? { parameterSpecs } : {}),

src/modules/llms/server/openai/models/chutesai.models.ts

@@ -115,15 +115,20 @@ export function chutesAIModelsToModelDescriptions(wireModels: unknown): ModelDes
     if (_chutesDenyListContains.some(contains => model.id.includes(contains)))
       continue;
 
+    // skip non-text-output models (image/video generation): they would masquerade as chat here
+    if (model.output_modalities?.length && !model.output_modalities.includes('text'))
+      continue;
+
     try {
       // label / description
       const label = _prettyModelId(model.id);
       const ownerStr = model.owned_by ? serverCapitalizeFirstLetter(model.owned_by) : '';
       const quantStr = model.quantization ? ` (${model.quantization})` : '';
       const description = `${ownerStr || 'Model'}${quantStr} via ChutesAI.`;
 
-      // context window: prefer explicit `context_length`, fall back to legacy `max_model_len`, then 8K
-      const contextWindow = model.context_length || model.max_model_len || 8192;
+      // context window: prefer explicit `context_length`, legacy `max_model_len`, else null (never a guess)
+      // no '[?]' marker (evaluated 2026-08-14): API-characterized (output_modalities drop above) - see llmsLabelUncurated
+      const contextWindow = model.context_length || model.max_model_len || null;
       const maxCompletionTokens = model.max_output_length || undefined;
 
       // interfaces: derive from actual fields rather than assume

src/modules/llms/server/openai/models/deepseek.models.ts

@@ -2,7 +2,7 @@ import { LLM_IF_HOTFIX_StripImages, LLM_IF_OAI_Chat, LLM_IF_OAI_Fn, LLM_IF_OAI_R
 
 import type { ModelDescriptionSchema } from '../../llm.server.types';
 
-import { llmsDefineManualMappings, fromManualMapping } from '../../models.mappings';
+import { llmsDefineManualMappings, fromManualMapping, llmsLabelUncurated } from '../../models.mappings';
 
 // --- Deepseek Model ID inference (auto-derived from _knownDeepseekChatModels) ---
 export type LlmsDeepseekModelId = typeof _knownDeepseekChatModels[number]['idPrefix'];
@@ -123,10 +123,9 @@ export function deepseekModelFilter(deepseekModelId: string) {
 export function deepseekModelToModelDescription(deepseekModelId: string): ModelDescriptionSchema {
   return fromManualMapping(_knownDeepseekChatModels, deepseekModelId, undefined, undefined, {
     idPrefix: deepseekModelId,
-    label: deepseekModelId.replaceAll(/[_-]/g, ' '),
-    description: 'New Deepseek Model',
-    contextWindow: 128000,
-    maxCompletionTokens: 4096,
+    label: llmsLabelUncurated(deepseekModelId.replaceAll(/[_-]/g, ' ')),
+    description: 'New DeepSeek arrival, not yet curated - capabilities and context window unverified.',
+    contextWindow: null,
     interfaces: [LLM_IF_OAI_Chat],
     hidden: true,
   });

src/modules/llms/server/openai/models/groq.models.ts

@@ -2,7 +2,7 @@ import { LLM_IF_OAI_Chat, LLM_IF_OAI_Fn, LLM_IF_OAI_Reasoning, LLM_IF_OAI_Vision
 import { Release } from '~/common/app.release';
 
 import type { ModelDescriptionSchema } from '../../llm.server.types';
-import { type KnownLink, type KnownModel, formatPubDate, fromManualMapping, llmDevCheckModels_DEV, llmsDefineModels } from '../../models.mappings';
+import { type KnownLink, type KnownModel, formatPubDate, fromManualMapping, llmDevCheckModels_DEV, llmsDefineModels, llmsLabelUncurated } from '../../models.mappings';
 
 // --- Groq Model ID inference (auto-derived from _knownGroqModels) ---
 export type LlmsGroqModelId = typeof _knownGroqModels[number]['idPrefix'];
@@ -182,9 +182,9 @@ export function groqModelToModelDescription(_model: unknown): ModelDescriptionSc
 
   const description = fromManualMapping(_knownGroqModels, model.id, model.created, undefined, {
     idPrefix: model.id,
-    label: model.id.replaceAll(/[_-]/g, ' '),
-    description: 'New Model',
-    contextWindow: model.context_window || 32768,
+    label: llmsLabelUncurated(model.id.replaceAll(/[_-]/g, ' ')),
+    description: 'New Groq arrival, not yet curated - capabilities unverified.',
+    contextWindow: model.context_window || null, // API value when present; null (not a guess) otherwise
     interfaces: [LLM_IF_OAI_Chat],
     hidden: true,
   });

src/modules/llms/server/openai/models/mistral.models.ts

@@ -221,7 +221,7 @@ export function mistralModels(wireModels: unknown): ModelDescriptionSchema[] {
       created: created || 0,
       updated: /*updated ||*/ created || 0,
       description: description,
-      contextWindow: max_context_length ?? 32768,
+      contextWindow: max_context_length ?? null, // schema requires the field today; null (not a guess) if it ever loosens
       interfaces: _mistralCapabilitiesToInterfaces(capabilities, id),
       // parameterSpecs: ...
       // maxCompletionTokens: ...

src/modules/llms/server/openai/models/moonshot.models.ts

@@ -3,7 +3,7 @@ import * as z from 'zod/v4';
 import { LLM_IF_HOTFIX_NoTemperature, LLM_IF_HOTFIX_StripImages, LLM_IF_OAI_Chat, LLM_IF_OAI_Fn, LLM_IF_OAI_Json, LLM_IF_OAI_PromptCaching, LLM_IF_OAI_Reasoning, LLM_IF_OAI_Vision } from '~/common/stores/llms/llms.types';
 
 import type { ModelDescriptionSchema } from '../../llm.server.types';
-import { llmsDefineModels, fromManualMapping, KnownModel } from '../../models.mappings';
+import { llmsDefineModels, fromManualMapping, KnownModel, llmsLabelUncurated } from '../../models.mappings';
 
 // --- Moonshot Model ID inference (auto-derived from _knownMoonshotModels) ---
 export type LlmsMoonshotModelId = typeof _knownMoonshotModels[number]['idPrefix'];
@@ -335,9 +335,10 @@ export function moonshotModelToModelDescription(_model: unknown): ModelDescripti
   const description = fromManualMapping(_knownMoonshotModels, model.id, model.created, undefined, {
     // NOTE: default: let us know if any of these show up
     idPrefix: model.id,
-    label: model.id.replaceAll(/[_-]/g, ' '),
-    description: 'Unknown Moonshot Model',
-    contextWindow: model.context_length || 128000,
+    label: llmsLabelUncurated(model.id.replaceAll(/[_-]/g, ' ')),
+    description: 'New Moonshot arrival, not yet curated - capabilities unverified.',
+    contextWindow: model.context_length || null, // API value when present; null (not a guess) otherwise
+    // optimistic capability leeway for 0-day arrivals; rein in when cataloged
     interfaces: [LLM_IF_OAI_Chat, LLM_IF_OAI_Fn, LLM_IF_OAI_Json, LLM_IF_OAI_PromptCaching],
     hidden: model.id.startsWith('moonshot-'), // hide older
   });

src/modules/llms/server/openai/models/novita.models.ts

@@ -75,8 +75,9 @@ export function novitaModelsToModelDescriptions(wireModels: unknown): ModelDescr
       if (description.length > 200)
         description = description.slice(0, 200) + '...';
 
-      // Context window
-      const contextWindow = model.context_size || 8192;
+      // Context window: API value or null, never a guess
+      // no '[?]' marker (evaluated 2026-08-14): API-characterized (model_type/endpoints filters) - see llmsLabelUncurated
+      const contextWindow = model.context_size || null;
 
       // Max completion tokens
       const maxCompletionTokens = model.max_output_tokens || undefined;