Skip to main content

Wafer Provider Guide

Wafer is a Tier-2 catalog provider: OpenAI-wire-compatible with no behavioural quirks, so its entire integration is one JSON file (src/lib/providers/catalog/wafer.json) rather than hand-written code. That file is the source of truth for everything on this page.

Verification status: this entry is docs- and roster-verified, not yet live-verified. The fields below come from Wafer's own public pages and unauthenticated GET requests — no account was created and no API key was used to build it. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

Key Facts​

  • Provider id: wafer
  • Protocol: OpenAI-compatible (/chat/completions) — the docs open with "Access Wafer models through an OpenAI-compatible API." (docs)
  • Base URL: https://pass.wafer.ai/v1
  • Authentication: the docs' Connect table shows Authorization: Bearer <YOUR_WAFER_API_KEY>
  • Default model: GLM-5.3
  • Models in catalog: 8 (the 8 models on the 2026-09-29 roster)
  • Streaming: true — the roster lists chat_completions.streaming: true on the 8 models
  • Tool calling: true — the roster lists chat_completions.tools: true on the 8 models
  • Tools while streaming: true — the roster lists chat_completions.tool_streaming: true on the 8 models
  • Structured output: true — the roster lists chat_completions.json_object: true and chat_completions.json_schema: true on the 8 models
  • Structured output + tools together: false — the roster lists chat_completions.tools_with_response_format: true, and no combined tools+schema request was sent, so the entry does not declare it
  • Embeddings: not declared (false)
  • Thinking: not declared (false) — the roster lists reasoning: true and a reasoning_effort block on the 8 models (see Models); no thinkingLevel request was sent
  • Billing policy field: no-free-tier — see Billing
  • Key format: none declared

Quick Start​

1. Get an API key​

  1. Visit: https://app.wafer.ai and sign in — the docs say "Sign in at app.wafer.ai to add credits and create an API key. Store your key securely." (https://docs.wafer.ai/serverless)
  2. Add credits and create an API key — the docs say "Add credits, create an API key, and pay per token." (https://docs.wafer.ai/serverless)
  3. Set WAFER_API_KEY in your .env file (the docs' Authentication row shows the Authorization: Bearer header)

2. Configure​

export WAFER_API_KEY=your-api-key
export WAFER_MODEL=GLM-5.3   # optional — overrides the default model
export WAFER_BASE_URL=https://pass.wafer.ai/v1   # optional — proxy or self-hosted gateway

3. Use it​

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "wafer",
  model: "GLM-5.3",
});

console.log(result.content);
# CLI
npx @juspay/neurolink generate "Hello" --provider wafer

Per-request credentials use the same shape as other NeuroLink providers:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "wafer",
  credentials: { wafer: { apiKey: process.env.WAFER_API_KEY } },
});

Billing​

The docs page https://docs.wafer.ai/serverless says: "Access Wafer models through an OpenAI-compatible API. Add credits, create an API key, and pay per token." Wafer's information page https://www.wafer.ai/ai-info says: "Hosted open models through an OpenAI-compatible API, with usage billed per token. Customers add credits and create an API key in the Wafer app."

The catalog's billingPolicy field takes one of free-tier, free-with-card or no-free-tier; the schema has no unknown value. The entry records no-free-tier as the schema's closest choice — it is not a Wafer statement.

Models​

ModelContextMax outputVision$/M in · out (cache read)Notes
GLM-5.3 ⭐1,048,576—no$0.29 / $4.40 ($0.2359)NeuroLink default; the model in the docs' curl request examples.
DeepSeek-V4-Pro1,048,576—no$0.40 / $4.20 ($0.32)Fallback.
Kimi-K31,048,576131,072yes$1.44 / $9.00 ($0.30)Fallback; roster capabilities.vision: true.
GLM-5.21,048,576—no$0.32 / $4.40 ($0.2599)Fallback.
Qwen3.8-27B262,144—yes$0.04 / $4.40 ($0.0342)Roster capabilities.vision: true; models.visionModel.
GLM-5.3-Flash1,048,576—yes$0.75 / $0.50 ($0.03)Roster capabilities.vision: true; roster description ends "Available serverless to gateway partners."
DeepSeek-V4.1-Flash1,048,576—yes$0.02 / $0.70 ($0.016)Roster capabilities.vision: true.
DeepSeek-V4-Flash-0731-Fast1,048,576—no$0.07 / $0.35 ($0.0531)Roster-listed.

Context, max output, prices and vision come from an unauthenticated GET https://pass.wafer.ai/v1/models made 2026-09-29 (8 models; the 8 are in this catalog). The roster gives prices as integer cents per million tokens (input_cents_per_million, output_cents_per_million), shown above divided by 100. The cache-read figure uses cache_read_microcents_per_million, which carries more digits than the cents field (for GLM-5.3: 23,590,000 microcents, shown as $0.2359; the cents field reads 24). The roster lists max_output_tokens: 131072 for Kimi-K3. The docs say "Use a model’s id in your requests." (docs)

models.defaultContextWindow (128,000) and models.defaultMaxOutputTokens (16,384) are placeholders the vendor does not publish; they apply to a model id outside the catalog.

Vision tags (vision: true) come from the roster's capabilities.vision and supports_vision fields, not from an image-input request — no key was used to send a real image.

Fallback order when the default is unavailable: DeepSeek-V4-Pro → Kimi-K3 → GLM-5.2. The runtime fallback model name the loader derives (fallbacks[1]) is Kimi-K3.

Roster descriptions​

The strings below are the wafer.description field of the roster entries, quoted as returned by GET https://pass.wafer.ai/v1/models on 2026-09-29.

  • GLM-5.3: "GLM-5.3 — Z.ai's flagship GLM-5.3 MoE (391B, NVFP4), self-hosted on the Wafer fleet with a 1M-token context window and frontier coding/agentic performance. Available serverless."
  • DeepSeek-V4-Pro: "DeepSeek-V4-Pro -- DeepSeek's V4 Pro MoE, self-hosted on the Wafer fleet with a 1M-token context window, reasoning-effort control and tool calling. Available serverless."
  • Kimi-K3: "Kimi K3 sparse MoE model, self-hosted on the Wafer fleet. Available serverless with a 1M-token context window and strong coding/agentic performance."
  • GLM-5.2: "GLM-5.2 — Z.ai's GLM-5.2 MoE (NVFP4), self-hosted on the Wafer fleet with a 1M-token context window, strong coding and reasoning, and tool calling. Available serverless."
  • Qwen3.8-27B: "Qwen3.8-27B -- Alibaba's dense 27B vision-language model, self-hosted on the Wafer fleet with a 262K-token context window, image and video input, reasoning-effort control and tool calling. Available serverless."
  • GLM-5.3-Flash: "GLM-5.3-Flash — Z.ai's fast GLM-5.3 MoE variant, self-hosted on the Wafer fleet with a 1M-token context window and strong coding/agentic performance. Available serverless to gateway partners."
  • DeepSeek-V4.1-Flash: "DeepSeek-V4.1-Flash — DeepSeek's V4.1 Flash MoE (552B total, 8B active for prefill / 16B for decode, 1M-token context), self-hosted on the Wafer fleet with reasoning-effort control and tool calling. Available serverless."
  • DeepSeek-V4-Flash-0731-Fast: "The same model served for high TPS."

Roster reasoning_effort blocks​

The roster lists a reasoning_effort block on the 8 models (control, efforts, default). These are roster values; no request set an effort.

Modelcontroleffortsdefault
GLM-5.3, GLM-5.2, GLM-5.3-Flash, DeepSeek-V4-Pro, DeepSeek-V4-Flash-0731-Fastsystem_promptnone, low, high, maxlow
Kimi-K3native_effortnone, high, maxnone
Qwen3.8-27Bnative_effortnone, low, medium, high, maxhigh
DeepSeek-V4.1-Flashnative_effortnone, low, high, maxnone

Verification status​

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog currently records for Wafer — docs- and roster-verified, not live-verified:

ProbeResult
Rosterunauthenticated GET https://pass.wafer.ai/v1/models, HTTP 200, 8 models, 2026-09-29 — no key used
Billinghttps://docs.wafer.ai/serverless: "Add credits, create an API key, and pay per token." — recorded as no-free-tier, the schema's closest value (2026-09-29)
Auth-failure shapeunauthenticated GET https://pass.wafer.ai/v1/chat/completions returned HTTP 401 with error.type authentication_error and error.code missing_api_key, 2026-09-29 (a GET; no POST was sent) — errorRules match status 401
Tools / structured outputroster chat_completions flags on the 8 models: tools, tool_streaming, json_object, json_schema and tools_with_response_format are true. These flags were not exercised with a request, and no combined tools+schema request was sent — structuredOutputWithTools stays false
Live capability sweepnot run. evidence.liveMatrix is null. Before treating this provider as production-ready, run npx tsx test/continuous-test-suite-provider-matrix.ts --provider=wafer with a real key and record the result.

Pages opened for this entry: https://www.wafer.ai/ai-info, https://docs.wafer.ai/serverless, https://docs.wafer.ai/serverless/zero-data-retention, https://www.wafer.ai/terms and https://app.wafer.ai.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

Troubleshooting​

SymptomCauseFix
HTTP 401 missing_api_keyWafer's response body reads: Missing API key. Provide credentials via Authorization: Bearer <key> or x-api-key: <key>.Set WAFER_API_KEY; NeuroLink's Tier-2 wire format sends Authorization: Bearer, one of the two headers Wafer's message names
Structured output ignored with tools attachedstructuredOutputWithTools is false on this entry — no combined request was sentNeuroLink omits response_format automatically whenever tools are present, before sending

See also​