@juspay/neurolink
Version:
The pipe layer of an AI nervous system: one interface connecting provider neurons to your application, across three inference types — generate, stream and decide. `decide` returns typed, calibrated judgements from a non-generative model (~400ms, ~$0.00002
115 lines (114 loc) • 9.31 kB
JSON
{
"$schema": "./provider-catalog.schema.json",
"id": "novita",
"displayName": "Novita AI",
"aliases": [],
"tier": 2,
"wire": {
"baseURL": "https://api.novita.ai/openai/v1"
},
"models": {
"default": "zai-org/glm-5.3-flash",
"testModel": "zai-org/glm-5.3-flash",
"fallbacks": [
"meta-llama/llama-3.3-70b-instruct",
"deepseek/deepseek-v3-0324",
"qwen/qwen-2.5-72b-instruct",
"google/gemma-3-27b-it"
],
"defaultContextWindow": 128000,
"defaultMaxOutputTokens": 4096,
"catalog": {
"meta-llama/llama-3.3-70b-instruct": {
"vision": false,
"status": "production",
"description": "Meta Llama 3.3 70B Instruct; a plain non-reasoning chat model. Kept as a fallback for plain chat because it does not support structured output; the runtime/default model is zai-org/glm-5.3-flash (see evidence.liveMatrix)"
},
"zai-org/glm-5.3-flash": {
"pricingPerMTok": {
"input": 0.075,
"output": 0.25,
"cachedInput": 0.015
},
"vision": false,
"status": "production",
"description": "Zhipu GLM-5.3 Flash \u2014 Novita's own served default and the model exercised end-to-end by NeuroLink's live capability probe (chat, streaming, tool calling, tools+streaming, tools+schema together). Reasons internally via `reasoning_content` before answering, so it needs a generous max_tokens budget or replies get cut mid-thought \u2014 pin `testModel` here for the capability matrix. Pricing from the authenticated roster, 2026-09-05"
},
"deepseek/deepseek-v3-0324": {
"vision": false,
"status": "production",
"description": "DeepSeek V3 (0324 snapshot) \u2014 the non-reasoning DeepSeek chat line (the reasoning variant is deepseek/deepseek-r1); fallback"
},
"qwen/qwen-2.5-72b-instruct": {
"vision": false,
"status": "production",
"description": "Qwen 2.5 72B Instruct \u2014 plain (non-thinking) Alibaba Qwen chat model; fallback"
},
"google/gemma-3-27b-it": {
"vision": false,
"status": "production",
"description": "Google Gemma 3 27B IT \u2014 non-reasoning instruct chat model; fallback"
}
},
"topModels": [
"meta-llama/llama-3.3-70b-instruct",
"zai-org/glm-5.3-flash",
"deepseek/deepseek-v3-0324",
"qwen/qwen-2.5-72b-instruct",
"google/gemma-3-27b-it"
]
},
"capabilities": {
"text": true,
"streaming": true,
"tools": true,
"toolsWithStreaming": true,
"structuredOutput": true,
"structuredOutputWithTools": true,
"embeddings": false,
"thinking": false
},
"errorRules": [
{
"status": 401,
"pattern": "failed to authenticate API key|FAILED_TO_AUTH",
"class": "authentication",
"message": "Invalid Novita API key. Check {apiKeyEnvVar}. Get one at https://novita.ai/settings/key-management"
},
{
"status": 404,
"pattern": "model not found|MODEL_NOT_FOUND",
"class": "invalid-model",
"message": "Novita model '{model}' not found. Pick a current model from the authenticated /openai/v1/models roster or https://novita.ai/models"
}
],
"setup": {
"url": "https://novita.ai/settings/key-management",
"apiKeyFormat": "^sk_[A-Za-z0-9_-]+$",
"billingPolicy": "free-tier",
"instructions": [
"1. Visit: https://novita.ai/settings/key-management and sign in (Google, GitHub, Hugging Face, or email)",
"2. Choose Create API Key, name it, and copy it immediately \u2014 it is shown in full only once and starts with sk_",
"3. New accounts start on a small trial voucher (a few dollars of Model API credit); manual top-ups go through Stripe with a $10 minimum, so keep early test calls small",
"4. Set {apiKeyEnvVar} in your .env file"
],
"description": "OpenAI-compatible endpoint at https://api.novita.ai/openai/v1. New accounts start on a small trial voucher; the roster spans 150+ open-weight chat, reasoning and vision-named models."
},
"evidence": {
"rosterVerified": {
"date": "2026-09-05",
"method": "authenticated GET /openai/v1/models \u2014 156 ids, e.g. zai-org/glm-5.3-flash, zai-org/glm-5.3, deepseek/deepseek-v4-pro-0813, moonshotai/kimi-k3",
"status": 200
},
"authProbe": {
"date": "2026-09-05",
"status": 401,
"code": "FAILED_TO_AUTH"
},
"liveMatrix": {
"date": "2026-09-05",
"result": "18-probe harness against zai-org/glm-5.3-flash (Novita's own served default) via https://api.novita.ai/openai/v1. Wire mechanics: chat, system role, array content parts, sampling params and max_completion_tokens all 200; SSE streaming 200 with a usage chunk and a terminal [DONE]. Reasoning is genuinely heavy on this model \u2014 several 200 responses show finish_reason=length with only 8 completion tokens, entirely spent on an unfinished reasoning_content and an empty visible content \u2014 but `default` and `testModel` are both pinned to the probed model itself (zai-org/glm-5.3-flash, Novita's own served default), which is exactly what the capability matrix below exercises — see the structured-output findings further down for why the reasoning-heavy default is still the right choice over the plain meta-llama/llama-3.3-70b-instruct fallback. Tool calling: a real tool_call (get_time, parsed tz argument) on the first turn, a null-content assistant tool_calls turn accepted on replay, and a tool_call surfacing mid-SSE-stream \u2014 tools and toolsWithStreaming true; tool_choice is tolerated even with no tools declared. response_format json_object returns 200 cleanly \u2014 structuredOutput true. response_format json_schema also returns 200, but at this token budget the content is unfinished reasoning text rather than schema-conformant JSON, same reasoning-budget effect as above \u2014 NeuroLink's coerceJsonToSchema repair layer is the safety net here, as for other reasoning-heavy vendors. Sending tools and a JSON schema together in one request succeeds (200, real tool_call returned) at the wire level. That single-shot result alone does not demonstrate the combined capability \u2014 the follow-up turn where the model must answer the original JSON-schema question from a tool result was untested at raw-probe time. It was proved separately through the SDK's public surface: `generate({ provider: 'novita', schema, ...})` with a registered getTime tool called end-to-end via dist/index.js \u2014 the model invoked getTime(tz='Asia/Tokyo') (toolsUsed=['getTime'], stepsUsed=2), then answered `{\"city\":\"Tokyo\",\"time\":\"09:00\"}`, which parsed against the request schema with no repair needed (jsonRepaired/jsonTruncated both unset) \u2014 passed 1/1 in the final live proof. structuredOutputWithTools is declared true on that basis. Vision: a 64x64 PNG sent to glm-5.3-flash (not a vision-tagged model; Novita has no live-verified vision model) returns 200 with reasoning_content correctly describing 'a solid blue rectangle' but an EMPTY final content \u2014 not a working, schema-usable vision reply, so no catalog entry is marked vision:true and models.visionModel is left unset. bad key -> 401 {\"code\":401,\"reason\":\"FAILED_TO_AUTH\",\"message\":\"failed to authenticate API key\"}; bad model -> 404 {\"code\":404,\"reason\":\"MODEL_NOT_FOUND\",\"message\":\"model not found\"}. Novita funds this account with a $1 Model API voucher, so only this one model was live-probed end-to-end; the other four catalog entries are roster-verified by id only, not capability-probed. Structured output on Novita is per-model, and the catalog schema can only express it per provider. Verified 2026-09-05 at the wire: meta-llama/llama-3.3-70b-instruct returns HTTP 400 for response_format json_schema (\"does not support 'json_schema' response format. Supported formats: json_object\") and also 400 for json_object, while zai-org/glm-5.3-flash returns 200 for json_object, json_schema and tools+json_schema together. The default is therefore zai-org/glm-5.3-flash, which is also Novita's own served default, so the provider-level structuredOutput and structuredOutputWithTools flags are true for the model a default-configuration caller actually gets. meta-llama/llama-3.3-70b-instruct remains a fallback for plain chat. Tool calling was verified separately on both models (finish_reason=tool_calls with a structured call). Note that glm-5.3-flash is a reasoning model: at very small max_tokens (observed at 40) the budget is consumed by reasoning and content comes back empty; the entry's defaultMaxOutputTokens of 4096 is well clear of that."
},
"addedInPR": "pending"
}
}