@juspay/neurolink 12.38.1 → 12.39.0

package/dist/providers/catalog/inference-net.json

@@ -0,0 +1,172 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "inference-net",
+    "displayName": "Inference.net",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://api.inference.net/v1",
+        "envOverrides": {
+            "apiKey": "INFERENCE_API_KEY"
+        }
+    },
+    "models": {
+        "default": "glm-5.2",
+        "fallbacks": [
+            "glm-5.3",
+            "glm-5.3-flash",
+            "llama-3.3-70b-instruct",
+            "deepseek-v3.2",
+            "gpt-4.1-mini",
+            "gemini-2.5-flash",
+            "claude-haiku-4-5"
+        ],
+        "defaultContextWindow": 1000000,
+        "defaultMaxOutputTokens": 128000,
+        "catalog": {
+            "glm-5.2": {
+                "contextWindow": 1000000,
+                "maxOutputTokens": 128000,
+                "pricingPerMTok": {
+                    "input": 1.3,
+                    "output": 4.4,
+                    "cachedInput": 0.2
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Zhipu GLM-5.2, served natively on Inference.net's own infrastructure. This is the exact model used in the request-body example on the vendor's own API quickstart (https://docs.inference.net/api/api-quickstart, retrieved 2026-09-28: `baseURL: \"https://api.inference.net/v1\"`, `model: \"glm-5.2\"`, streaming chat via the OpenAI SDK), so it is picked as the catalog default over an arbitrary roster pick. context_length/max_completion_tokens/pricing read from the live, unauthenticated GET https://api.inference.net/v1/models on 2026-09-28 (200, 63 models). input_modalities is [\"text\"] only for this model — no image/video support documented. Not live capability-probed (credential-free onboarding, no key used)."
+            },
+            "glm-5.3": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 131072,
+                "pricingPerMTok": {
+                    "input": 0.9,
+                    "output": 3.0,
+                    "cachedInput": 0.15
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Zhipu GLM-5.3 — newer GLM generation on the same live roster; one of only a handful of models whose supported_endpoints spans chat, responses and messages together (per the live /v1/models payload, 2026-09-28). input_modalities text-only. Roster-listed and priced from the vendor's own /v1/models response; not itself capability-probed."
+            },
+            "glm-5.3-flash": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 128000,
+                "pricingPerMTok": {
+                    "input": 0.09,
+                    "output": 0.28,
+                    "cachedInput": 0.02
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Zhipu GLM-5.3 Flash — cheapest model on the roster and the only one of the eight cataloged here whose live /v1/models entry lists input_modalities [\"text\",\"image\",\"video\"] (2026-09-28 GET), so it is the one model in this entry marked vision: true; not itself capability-probed end to end (no key used)."
+            },
+            "llama-3.3-70b-instruct": {
+                "contextWindow": 131072,
+                "maxOutputTokens": 128000,
+                "pricingPerMTok": {
+                    "input": 0.13,
+                    "output": 0.4
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Meta Llama 3.3 70B Instruct, open-weight — widely-recognized baseline fallback. Roster-listed with no cached-input price on the live /v1/models payload (2026-09-28); input_modalities text-only."
+            },
+            "deepseek-v3.2": {
+                "contextWindow": 163840,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.7,
+                    "output": 2.0,
+                    "cachedInput": 0.35
+                },
+                "vision": false,
+                "status": "production",
+                "description": "DeepSeek V3.2, open-weight — roster-listed fallback from the live /v1/models GET (2026-09-28); input_modalities text-only."
+            },
+            "gpt-4.1-mini": {
+                "contextWindow": 1047576,
+                "maxOutputTokens": 32768,
+                "pricingPerMTok": {
+                    "input": 0.4,
+                    "output": 1.6,
+                    "cachedInput": 0.1
+                },
+                "vision": true,
+                "status": "production",
+                "description": "OpenAI GPT-4.1 mini, reachable through Inference.net's hosted gateway under the same Bearer-key API (the vendor's quickstart separately documents a 'proxy mode' with extra x-inference-provider* headers for BYO provider keys — not used here; this catalog entry only relies on the plain gateway call). Live /v1/models (2026-09-28) lists input_modalities [\"text\",\"image\"], so marked vision: true; not capability-probed end to end."
+            },
+            "gemini-2.5-flash": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.3,
+                    "output": 2.5,
+                    "cachedInput": 0.03
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Google Gemini 2.5 Flash, reachable through the same gateway. Live /v1/models (2026-09-28) lists input_modalities [\"text\",\"image\"], so marked vision: true; not capability-probed end to end."
+            },
+            "claude-haiku-4-5": {
+                "contextWindow": 200000,
+                "maxOutputTokens": 64000,
+                "pricingPerMTok": {
+                    "input": 1.0,
+                    "output": 5.0,
+                    "cachedInput": 0.1
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Anthropic Claude Haiku 4.5, reachable through the same gateway; the live /v1/models payload notes Claude-family entries list only the \"messages\" supported_endpoint (2026-09-28). input_modalities [\"text\",\"image\"], so marked vision: true; not capability-probed end to end."
+            }
+        },
+        "topModels": [
+            "glm-5.2",
+            "glm-5.3",
+            "glm-5.3-flash",
+            "llama-3.3-70b-instruct",
+            "deepseek-v3.2",
+            "gpt-4.1-mini",
+            "gemini-2.5-flash",
+            "claude-haiku-4-5"
+        ]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": true,
+        "toolsWithStreaming": false,
+        "structuredOutput": true,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": false
+    },
+    "errorRules": [
+        {
+            "status": 401,
+            "class": "authentication",
+            "message": "Invalid Inference.net API key. Check {apiKeyEnvVar}. Get one at {setupUrl}"
+        }
+    ],
+    "setup": {
+        "url": "https://inference.net/register/",
+        "apiKeyFormat": null,
+        "billingPolicy": "free-tier",
+        "instructions": [
+            "1. Visit: https://inference.net/register/ and sign up",
+            "2. Create an Inference.net project API key from the dashboard (this is the gateway key that goes in {apiKeyEnvVar} — do not confuse it with a third-party provider key used only for the separate 'proxy mode')",
+            "3. New accounts start on a pay-as-you-go, $0-usage tier (1M included gateway requests/month, 30 req/min) per https://inference.net/pricing; confirm current terms before heavy use",
+            "4. Set {apiKeyEnvVar} in your .env file"
+        ],
+        "description": "OpenAI-compatible endpoint at https://api.inference.net/v1 (`Authorization: Bearer <key>`), serving 63 models on the live roster across many upstream families — Claude, GPT, Gemini, DeepSeek, GLM, Kimi, Llama, Qwen, Gemma, Nemotron, plus Inference.net's own Schematron structured-extraction models — behind one API. A separate 'proxy mode' can forward to your own provider key via x-inference-provider* headers; not used by this catalog entry."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://api.inference.net/v1/models — 63 models; credential-free, no key used",
+            "status": 200
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1848"
+    }
+}

package/dist/providers/catalog/koscompute.json

@@ -0,0 +1,81 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "koscompute",
+    "displayName": "Kosmik Compute",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://api.koscompute.com/v1"
+    },
+    "models": {
+        "default": "qwen/qwen3.8-27b",
+        "fallbacks": ["qwen/qwen3.8-27b"],
+        "defaultContextWindow": 262144,
+        "defaultMaxOutputTokens": 32768,
+        "catalog": {
+            "qwen/qwen3.8-27b": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 32768,
+                "pricingPerMTok": {
+                    "input": 0.35,
+                    "output": 2.2,
+                    "cachedInput": 0.09
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Qwen 3.8 27B (fp8) — the only chat-capable model in KosCompute's roster; the other 6 ids returned by /v1/models are Whisper transcription and TTS models (endpoint_kind audio/tts), not chat. Confirmed via unauthenticated GET https://api.koscompute.com/v1/models (2026-09-28, HTTP 200): context_length 262144 (modality_context_lengths.text), 131072 for image input; max_output_length/max_output 32768; pricing input $0.35 / output $2.20 per MTok, cached input $0.09 (the response's own precomputed input/output/cache_read fields, matching prompt/completion/input_cache_read per-token rates); supported_features [tool_calling, json_mode, structured_outputs, reasoning]; supported_endpoints [/v1/chat/completions, /v1/responses, /v1/messages]; input_modalities [text, image] with vision.max_pixels 1048576. model_info reports reasoning:true, extended_thinking:true, thinking_default:true, max_thinking_tokens:8192, supported_reasoning_efforts [none, low, medium, xhigh]. Docs corroborate generically for the chat-completions surface: api.koscompute.com/docs/guides/tool-calling/ (tools array, tool_choice auto|none|required|named, results read from choices[0].message.tool_calls), /docs/guides/structured-output/ (response_format json_object and json_schema, 'Unsupported models return a 400 capability error' for models lacking json_mode/structured_outputs — this model advertises both), and /docs/guides/streaming/ (stream:true, text/event-stream, data: [DONE]). No live capability probe was run (credential-free onboarding, no API key) — see evidence.liveMatrix: null."
+            }
+        },
+        "topModels": ["qwen/qwen3.8-27b"]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": true,
+        "toolsWithStreaming": true,
+        "structuredOutput": true,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": true
+    },
+    "errorRules": [
+        {
+            "status": 401,
+            "class": "authentication",
+            "message": "Invalid KosCompute API key. Check {apiKeyEnvVar}. See {setupUrl} for what KosCompute's docs publish on authentication — the vendor documents 401 as 'missing, invalid, expired, inactive, or conflicting credentials' but gives no example error body or code string."
+        },
+        {
+            "status": 404,
+            "pattern": "model_not_found",
+            "class": "invalid-model",
+            "message": "KosCompute model '{model}' not found. Pick a current id from an unauthenticated GET https://api.koscompute.com/v1/models."
+        },
+        {
+            "status": 429,
+            "pattern": "concurrent_limit_exceeded",
+            "class": "rate-limit",
+            "message": "KosCompute per-key concurrency limit exceeded for '{model}'. Wait for an in-flight request to finish, or reduce parallel requests, before retrying."
+        }
+    ],
+    "setup": {
+        "url": "https://api.koscompute.com/docs/overview/quickstart/",
+        "apiKeyFormat": null,
+        "billingPolicy": "no-free-tier",
+        "instructions": [
+            "1. KosCompute publishes no public signup, dashboard, or key-management page as of 2026-09-28 — the docs' quickstart (this URL) assumes you already hold a key and only shows how to use one, via `export KOSCOMPUTE_API_KEY=\"your-api-key\"`; the site's only 'Start building' call-to-action also points here, not to a signup form",
+            "2. The docs reference an account 'balance' (/docs/overview/authentication/: 'Authentication success does not guarantee model access, remaining account balance, or available capacity') but publish no pricing, free-tier, or top-up page, so billing policy could not be confirmed from public sources — treat this entry's `no-free-tier` as the conservative default in the absence of any documented free allocation, not a confirmed vendor policy",
+            "3. Once you have a key (provisioned out of band), set {apiKeyEnvVar} in your .env file",
+            "4. Sanity-check connectivity first with an unauthenticated GET https://api.koscompute.com/v1/models (no key needed) before spending a real chat request"
+        ],
+        "description": "OpenAI-compatible endpoint at https://api.koscompute.com/v1, described by the vendor as EU-hosted inference (docs: 'EU-hosted, OpenAI-compatible inference for chat, reasoning, vision, speech recognition, and speech synthesis'); also exposes Responses-API and Anthropic-Messages-compatible surfaces for the same chat model. No public signup or key-management page is published, and billing policy is unconfirmed."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://api.koscompute.com/v1/models — 7 models; credential-free, no key used",
+            "status": 200
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1848"
+    }
+}

package/dist/providers/catalog/lilac.json

@@ -0,0 +1,107 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "lilac",
+    "displayName": "Lilac",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://api.getlilac.com/v1"
+    },
+    "models": {
+        "default": "moonshotai/kimi-k2.6",
+        "fallbacks": [
+            "minimaxai/minimax-m3",
+            "zai-org/glm-5.2",
+            "google/gemma-4-31b-it"
+        ],
+        "defaultContextWindow": 262144,
+        "defaultMaxOutputTokens": 262144,
+        "catalog": {
+            "moonshotai/kimi-k2.6": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 262144,
+                "pricingPerMTok": {
+                    "input": 0.7,
+                    "output": 3.5,
+                    "cachedInput": 0.16
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Moonshot AI Kimi K2.6 — the model used in every code sample on Lilac's own quickstart and chat-completions docs (including its vision example), so it is chosen as `default` on that basis; no capability probe was possible without a key. Vendor description: 'Moonshot AI's multimodal agentic model for long-horizon coding, tool use, and design workflows.' Live roster (unauthenticated GET /v1/models, 2026-09-28): context_length 262144, top_provider.max_completion_tokens 262144 (not documented separately as a distinct output ceiling — the roster reports it equal to the context window, so treat it as an upper bound rather than a proven lower cap), INT4 quantization, input_modalities text+image (no video). supported_parameters includes response_format, structured_outputs, tools and tool_choice; supported_features is [tools, reasoning]. Pricing per https://docs.getlilac.com/inference/models (retrieved 2026-09-28, matches the live roster exactly): $0.70 / $3.50 per MTok input/output, $0.16 per MTok cached input. docs.getlilac.com/inference/openai-compatibility.md lists vision, tool/function calling and structured output (JSON mode) as generally-supported features of the API; this model's own architecture also reports image input, so vision is true here."
+            },
+            "minimaxai/minimax-m3": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 1048576,
+                "pricingPerMTok": {
+                    "input": 0.28,
+                    "output": 1.1,
+                    "cachedInput": 0.05
+                },
+                "vision": true,
+                "status": "production",
+                "description": "MiniMax M3 — the largest-context model on the roster. Vendor description: 'MiniMax's native multimodal 428B MoE reasoning model with a 1M-token context window and agentic tool-use support.' Live roster (2026-09-28): context_length 1048576, max_completion_tokens 1048576, FP8, input_modalities text+image+video (vision true). supported_parameters includes response_format, structured_outputs, tools, tool_choice and chat_template_kwargs (MiniMax's per-model reasoning toggle is `thinking_mode`). Pricing per docs.getlilac.com/inference/models (2026-09-28, matches roster): $0.28 / $1.10 per MTok input/output, $0.05 per MTok cached input. Roster-verified only, not capability-probed without a key; first fallback."
+            },
+            "zai-org/glm-5.2": {
+                "contextWindow": 524288,
+                "maxOutputTokens": 524288,
+                "pricingPerMTok": {
+                    "input": 0.9,
+                    "output": 3,
+                    "cachedInput": 0.17
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Zhipu GLM 5.2. Vendor description: '743B MoE model for long-context reasoning, coding, and agentic tool-use workflows.' Live roster (2026-09-28): context_length 524288, max_completion_tokens 524288, FP8/NVFP4, input_modalities text-only (no image/video — vision false, the only non-vision model in this roster). supported_parameters includes response_format, structured_outputs, tools, tool_choice and reasoning_effort — this is the one model whose docs say `reasoning_effort` also works as a plain top-level field (OpenAI-style), not only inside `chat_template_kwargs`. Pricing per docs.getlilac.com/inference/models (2026-09-28, matches roster): $0.90 / $3.00 per MTok input/output, $0.17 per MTok cached input. Roster-verified only, not capability-probed without a key; fallback."
+            },
+            "google/gemma-4-31b-it": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 262144,
+                "pricingPerMTok": {
+                    "input": 0.11,
+                    "output": 0.35
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Google Gemma 4 31B instruct-tuned. Vendor description: 'Google's open 31B instruction-tuned model.' Live roster (2026-09-28): context_length 262144, max_completion_tokens 262144, FP8, input_modalities text+image+video (vision true). supported_parameters includes response_format and tools but, unlike the other three models on this roster, does NOT list structured_outputs — and docs.getlilac.com/inference/models carries an explicit caveat for this model: combining its reasoning parser with `enable_thinking: false` 'can silently disable xgrammar-backed structured output.' Pricing per docs (2026-09-28, matches roster): $0.11 / $0.35 per MTok input/output; no cached-input rate is published for this model (the live roster returns no `input_cache_read` field for it either). Roster-verified only, not capability-probed without a key; fallback."
+            }
+        },
+        "topModels": [
+            "moonshotai/kimi-k2.6",
+            "minimaxai/minimax-m3",
+            "zai-org/glm-5.2",
+            "google/gemma-4-31b-it"
+        ]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": true,
+        "toolsWithStreaming": false,
+        "structuredOutput": true,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": true
+    },
+    "errorRules": [],
+    "setup": {
+        "url": "https://console.getlilac.com",
+        "apiKeyFormat": null,
+        "billingPolicy": "no-free-tier",
+        "instructions": [
+            "1. Visit: https://console.getlilac.com and sign up with your email or a Google account",
+            "2. Create an organization, then go to API Keys in the dashboard and create a new key — it is shown only once, so copy it immediately",
+            "3. Go to Billing, click Add Credits and complete the Stripe checkout (as little as $5.00) — a positive credit balance is required before you can make any API request; there is no free tier or free credits",
+            "4. Set {apiKeyEnvVar} in your .env file"
+        ],
+        "description": "OpenAI-compatible endpoint at https://api.getlilac.com/v1 — docs.getlilac.com/inference/openai-compatibility.md states it 'provides an OpenAI-compatible API' and that pointing the official OpenAI SDK at it with a new base_url/api_key is sufficient. GET /v1/models is public and unauthenticated (confirmed live, 200, 4 models, 2026-09-28). Embeddings, audio, fine-tuning, the Assistants API, Batch API and file uploads are explicitly documented as not yet available. Billing is prepaid credit via Stripe with no minimum spend or contract once funded, but a positive balance is required before any request succeeds — no free tier. Rate limit: 200 requests/minute per organization; excess requests may get HTTP 429 with a Retry-After header. No self-service signup URL was found in the initial lead for this vendor, but docs.getlilac.com/inference/quickstart.md (fetched 2026-09-28) documents a normal self-serve console at console.getlilac.com — new-model requests still go through contact@getlilac.com, and enterprises wanting monthly invoicing instead of prepaid credits can also email that address."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://api.getlilac.com/v1/models — 4 models; credential-free, no key used",
+            "status": 200
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1848"
+    }
+}

package/dist/providers/catalog/llmtech.json

@@ -0,0 +1,63 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "llmtech",
+    "displayName": "LLM Tech",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://api.llmtech.eu/v1"
+    },
+    "models": {
+        "default": "nvidia/Qwen3.8-27B-NVFP4",
+        "fallbacks": ["nvidia/Qwen3.8-27B-NVFP4"],
+        "defaultContextWindow": 262144,
+        "defaultMaxOutputTokens": 32768,
+        "catalog": {
+            "nvidia/Qwen3.8-27B-NVFP4": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 32768,
+                "pricingPerMTok": {
+                    "input": 0.25,
+                    "output": 2.09,
+                    "cachedInput": 0.04
+                },
+                "vision": true,
+                "status": "production",
+                "description": "LLM Tech's sole hosted chat model (NVFP4-quantized Qwen3.8-27B on Blackwell), self-hosted on EU hardware (GPU node in Italy, edge in Nuremberg). Re-verified live via unauthenticated GET https://api.llmtech.eu/v1/models on 2026-09-28 (HTTP 200, exactly 1 model): context_length 262144, max_output/max_output_tokens 32768, pricing_per_m input $0.25 / output $2.09 / cache_read $0.04 per MTok, and a features array of [tools, json_mode, structured_outputs, logprobs, reasoning, vision] on the model object itself. All of this matches the vendor's own model page (https://llmtech.eu/models/qwen3.8-27b, retrieved 2026-09-28), which documents: an OpenAI-compatible client.chat.completions.create(...) quickstart against this baseURL; SSE streaming ('SSE, OpenAI format', stream=True in the sample); tool calling ('yes, with structured outputs'); real vision support ('yes, the vision tower is kept in BF16 and works' — image tokens count toward prompt_tokens and bill at the input rate, e.g. a 64x64 PNG costs 64 image tokens; images must go in a user message, not a system message; the page notes third-party listings calling this model text-only are wrong); reasoning control via enable_thinking/reasoning_effort request parameters; automatic prompt caching billed at the cache rate; and zero data retention (prompts/completions 'never used for training'). No response_format wire example or combined tools+response_format request is shown or was probed (see capabilities.structuredOutputWithTools). Not itself capability-probed end to end — credential-free onboarding, no API key used."
+            }
+        },
+        "topModels": ["nvidia/Qwen3.8-27B-NVFP4"]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": true,
+        "toolsWithStreaming": false,
+        "structuredOutput": true,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": true
+    },
+    "errorRules": [],
+    "setup": {
+        "url": "https://llmtech.eu/models/qwen3.8-27b",
+        "apiKeyFormat": null,
+        "billingPolicy": "no-free-tier",
+        "instructions": [
+            "1. LLM Tech publishes no public signup or self-service key page as of 2026-09-28 (self-service is described on the vendor's site as 'planned'); email artem@llmtech.eu to request a key — the account dashboard (https://llmtech.eu/cabinet/) confirms keys are issued by the company rather than self-created ('we hold no passwords and no accounts'; 'paste the API key we issued you')",
+            "2. Once you receive a key, https://llmtech.eu/cabinet/ shows current-month billing (totals, a daily usage chart, invoice line items) by pasting the key in — there is no separate account or login",
+            "3. Billing is pay-per-token USD with 'no subscription, no minimums' (input $0.25/M, output $2.09/M, cached input $0.04/M per https://llmtech.eu/models/qwen3.8-27b); no free tier is documented, so budget for paid usage from the first call",
+            "4. Set {apiKeyEnvVar} in your .env file"
+        ],
+        "description": "OpenAI-compatible endpoint at https://api.llmtech.eu/v1, self-hosted on EU hardware (GPU node in Italy, edge in Nuremberg), serving a single model — nvidia/Qwen3.8-27B-NVFP4 — confirmed live by an unauthenticated GET /v1/models. No public signup exists; keys are issued manually by the vendor via email. The vendor states zero data retention: prompts and completions are never used for training."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://api.llmtech.eu/v1/models — 1 model; credential-free, no key used",
+            "status": 200
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1848"
+    }
+}

package/dist/providers/catalog/moark.json

@@ -0,0 +1,94 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "moark",
+    "displayName": "Moark",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://api.moark.com/v1"
+    },
+    "models": {
+        "default": "Qwen3-8B",
+        "fallbacks": [
+            "Qwen3-4B",
+            "ERNIE-4.5-Turbo",
+            "GLM-5.3",
+            "kimi-k3",
+            "DeepSeek-V4-Pro-0813"
+        ],
+        "defaultContextWindow": 32000,
+        "defaultMaxOutputTokens": 4096,
+        "catalog": {
+            "Qwen3-8B": {
+                "vision": false,
+                "status": "production",
+                "description": "Recommended — roster-verified via unauthenticated GET https://moark.com/v1/models (2026-09-28, HTTP 200, 222 models). Chosen as `default` because Moark's own text-generation guide (https://moark.com/docs/products/apis/texts/text-generation, retrieved 2026-09-28) names Qwen3-8B and Qwen3-4B specifically as its suggested free-testing/validation models ('免费在线体验', quick-validation picks) — the only concrete, documented model recommendation on that page that also matches a currently-live roster id; the page's other named examples (Qwen2.5-72B-Instruct used in every code sample, kimi-k2-instruct and DeepSeek-R1 for long-text/reasoning, ERNIE-4.5-Turbo for multilingual) are not presented as free and, except ERNIE-4.5-Turbo, are no longer on the live roster. No credential-free capability probe was possible (chat/completions is POST-only, and this onboarding used unauthenticated GETs only): tool calling, structured output (guided_json/guided_choice) and SSE streaming are documented generically for the API but not confirmed against this specific model. contextWindow/maxOutputTokens are left unset on every model in this catalog — Moark's docs give only a generic '32K、128K 等' illustration (not tied to any model) and GET /v1/models returns only {id, object, created, owned_by}, no context_length or pricing fields."
+            },
+            "Qwen3-4B": {
+                "vision": false,
+                "status": "production",
+                "description": "Roster-verified only (unauthenticated GET https://moark.com/v1/models, 2026-09-28). Named alongside Qwen3-8B as a free testing/validation model on Moark's text-generation guide (https://moark.com/docs/products/apis/texts/text-generation, retrieved 2026-09-28) — smaller sibling of the default. Not itself capability-probed; fallback."
+            },
+            "ERNIE-4.5-Turbo": {
+                "vision": false,
+                "status": "production",
+                "description": "Roster-verified only (unauthenticated GET https://moark.com/v1/models, 2026-09-28). Baidu ERNIE 4.5 Turbo — named by Moark's text-generation guide (retrieved 2026-09-28) as its suggested model for multilingual use cases. Not itself capability-probed; fallback."
+            },
+            "GLM-5.3": {
+                "vision": false,
+                "status": "production",
+                "description": "Roster-verified only (unauthenticated GET https://moark.com/v1/models, 2026-09-28). Zhipu GLM-5.3 — present on the live roster; Moark's text-generation guide does not name it specifically (its worked examples use Qwen2.5-72B-Instruct instead, which is not on the current roster). Included as a general-purpose fallback on roster presence alone; not capability-probed."
+            },
+            "kimi-k3": {
+                "vision": false,
+                "status": "production",
+                "description": "Roster-verified only (unauthenticated GET https://moark.com/v1/models, 2026-09-28). Moonshot AI Kimi K3 — the closest live-roster match to Moark's docs-suggested 'long text' pick (the exact doc-named id, kimi-k2-instruct, is no longer on the roster; Kimi-K2.7-Code and Kimi-K2.6 are also present but code-branded/older). Not capability-probed; fallback."
+            },
+            "DeepSeek-V4-Pro-0813": {
+                "vision": false,
+                "status": "production",
+                "description": "Roster-verified only (unauthenticated GET https://moark.com/v1/models, 2026-09-28). DeepSeek V4 Pro — the closest live-roster match to Moark's docs-suggested 'specialized reasoning' pick (the exact doc-named id, DeepSeek-R1, is no longer on the roster). Not capability-probed; fallback."
+            }
+        },
+        "topModels": [
+            "Qwen3-8B",
+            "Qwen3-4B",
+            "ERNIE-4.5-Turbo",
+            "GLM-5.3",
+            "kimi-k3",
+            "DeepSeek-V4-Pro-0813"
+        ]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": "model-dependent",
+        "toolsWithStreaming": false,
+        "structuredOutput": false,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": false
+    },
+    "errorRules": [],
+    "setup": {
+        "url": "https://moark.com/dashboard/tokens",
+        "apiKeyFormat": null,
+        "billingPolicy": "no-free-tier",
+        "instructions": [
+            "1. Visit: https://moark.com and sign in — no separate signup page is documented, only workbench login",
+            "2. Go to 工作台 → 设置 → 访问令牌 (Workbench → Settings → Access Tokens) at https://moark.com/dashboard/tokens and create a token",
+            "3. Fund a prepaid resource package before making live calls (e.g. the all-model package, https://moark.com/serverless-api/order?package=1910) — billing is prepaid, charged per-call or per-token (https://moark.com/docs/products/apis, retrieved 2026-09-28); models can be tried free in the web console, but no formal free-tier API policy is documented, and package usage is non-refundable (https://moark.com/docs/billing/purchase, retrieved 2026-09-28)",
+            "4. Set {apiKeyEnvVar} in your .env file"
+        ],
+        "description": "OpenAI-compatible endpoint at https://api.moark.com/v1 — Moark's own docs state '模力方舟的 Serverless API 兼容开发者喜爱且社区流行的 OpenAI 风格 API' (https://moark.com/docs/products/apis/texts/text-generation, retrieved 2026-09-28), serving 222 models (unauthenticated GET /v1/models, 2026-09-28) spanning text, vision, image, video, audio, embedding and reranking — this entry covers only the OpenAI-compatible chat/completions surface. The vendor also documents vLLM-style guided_json/guided_choice structured-decoding parameters, distinct from the OpenAI response_format field NeuroLink's generic client sends, so structuredOutput is not claimed here without a live probe."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://moark.com/v1/models — 222 models; credential-free, no key used",
+            "status": 200
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1848"
+    }
+}