Skip to main content

Cerebras Provider Guide

The fastest hosted generation available (~3000 tokens/s on GPT-OSS 120B) via Cerebras' Wafer-Scale Engine — best for throughput-hungry workloads

Overview​

Cerebras serves open-weight models on its Wafer-Scale Engine (WSE), a single wafer-sized chip whose on-die memory bandwidth yields generation speeds an order of magnitude above GPU clouds. NeuroLink wraps api.cerebras.ai/v1 (OpenAI-compatible, zero-quirk Tier 2 catalog entry) so the standard generate / stream contract applies.

The roster below was verified against a live authenticated /v1/models on 2026-09-28 — Cerebras retires models aggressively: gemma-4-31b left the roster in September 2026, and earlier llama/qwen ids return 404:

  • gpt-oss-120b (default) — OpenAI's open-weight 120B reasoning model, ~3000 tok/s

The roster also lists qwen-3.8-27b, which the catalog does not carry yet: no chat call has verified it.

Key Facts​

  • Protocol: OpenAI-compatible (/v1/chat/completions)
  • Default base URL: https://api.cerebras.ai/v1
  • Default model: gpt-oss-120b
  • Context window: 65K tokens on the free tier, 131K on paid tiers. NeuroLink budgets context against the 65K free-tier floor — the account tier isn't knowable from the key, and compacting early on a paid tier is safe while overrunning a 65K window is not.
  • Max output: 32K free / 40K paid
  • Vision: No — gemma-4-31b, the only vision-capable model, is retired
  • Streaming: Supported
  • Tool calling: Supported (native)
  • Structured output: Supported — but not combined with tools in one request: the API rejects tools + response_format together with 400 wrong_api_format ("tools" is incompatible with "response_format"). NeuroLink handles this the same way as Groq: with tools active the schema is enforced post-hoc on the final text instead of on the wire.
  • Reasoning trace: gpt-oss-120b emits reasoning deltas before content — see Troubleshooting for the maxTokens implication.
  • Billing: no keyless free tier. Even the one-time $5 promotional credit requires saving a payment method ("you won't be charged now"). Pay-as-you-go starts at $10.
  • Pricing (per million tokens, checked 2026-08-27): gpt-oss-120b $0.35 in / $0.75 out.

Quick Start​

1. Get an API Key​

Sign up at https://cloud.cerebras.ai (Google OAuth works), claim the $5 free credit under Billing → Credits (a payment card must be saved — no charge is made), and create an API key (prefix csk-).

2. Configure Environment​

# Required
CEREBRAS_API_KEY=csk-...

# Optional: override the default model (default: gpt-oss-120b)
CEREBRAS_MODEL=gpt-oss-120b

# Optional: override the base URL
# CEREBRAS_BASE_URL=https://api.cerebras.ai/v1

3. Generate Your First Response​

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

const ai = new NeuroLink();

const result = await ai.generate({
  provider: "cerebras",
  input: { text: "What is the Cerebras wafer-scale engine?" },
});

console.log(result.content);

SDK Usage​

Basic Generation​

const result = await ai.generate({
  provider: "cerebras",
  input: { text: "Write a haiku about silicon wafers." },
});

Streaming​

const stream = await ai.stream({
  provider: "cerebras",
  input: { text: "Explain how B-trees work, step by step." },
});

for await (const chunk of stream.stream) {
  if ("content" in chunk && chunk.content) process.stdout.write(chunk.content);
}

Tool Calling​

const ai = new NeuroLink();
ai.registerTool("getTime", {
  name: "getTime",
  description: "Returns the current UTC time",
  inputSchema: { type: "object", properties: {} },
  execute: async () => ({ utc: new Date().toISOString() }),
});

const result = await ai.generate({
  provider: "cerebras",
  input: { text: "What is the current UTC time? Use the tool." },
  enabledToolNames: ["getTime"],
});

Structured Output​

import { z } from "zod";

const result = await ai.generate({
  provider: "cerebras",
  input: { text: "Name three fast chips as JSON." },
  schema: z.object({ chips: z.array(z.string()) }),
  maxTokens: 1000,
});

console.log(result.structuredData); // parsed, schema-shaped object

Combining schema with active tools works, but the schema is enforced post-hoc rather than on the wire (see Key Facts) — expect structuredData to be best-effort in that combination, exactly as with Groq.

Per-Call Credentials​

const result = await ai.generate({
  provider: "cerebras",
  input: { text: "..." },
  credentials: { cerebras: { apiKey: "user-key" } },
});

CLI Usage​

# Default model (gpt-oss-120b)
pnpm run cli generate "Quick question" --provider cerebras

# Explicit model
pnpm run cli generate "Hi" --provider cerebras --model gpt-oss-120b

# Streaming
pnpm run cli stream "Count to ten" --provider cerebras

# Loop / chat
pnpm run cli loop --provider cerebras

Provider Aliases​

AliasExample
cerebras--provider cerebras

Configuration Reference​

Environment VariableRequiredDefaultDescription
CEREBRAS_API_KEYYes—Cerebras API key
CEREBRAS_MODELNogpt-oss-120bDefault model
CEREBRAS_BASE_URLNohttps://api.cerebras.ai/v1Base URL

Feature Support Matrix​

Featuregpt-oss-120b
Text generationYes
StreamingYes
Tool callingYes
Structured outputYes
Structured output + toolsPost-hoc
VisionNo
EmbeddingsNo
Context window65K/131K

Troubleshooting​

"Invalid Cerebras API key"​

# Presence check without printing the secret (terminal capture, CI logs
# and shell transcripts retain echoed values):
test -n "$CEREBRAS_API_KEY" && echo "CEREBRAS_API_KEY is set" || echo "CEREBRAS_API_KEY is missing"
export CEREBRAS_API_KEY=csk-...

Get / rotate at https://cloud.cerebras.ai. A bad key returns 401 with "code": "wrong_api_key", which NeuroLink maps to this message.

402 payment_required on every call​

The account has no balance. Cerebras has no keyless free tier: open Billing → Credits → ADD CREDITS in the console, choose "Start with limited free credits", and save a payment card — the $5 promo credit activates with no charge. Skipping the claim step during onboarding ("SKIP TO CONSOLE") leaves the balance at $0.00.

Empty content with small maxTokens on gpt-oss-120b​

gpt-oss-120b is a reasoning model: it spends its first tokens on a reasoning channel before emitting content. With a tight budget (e.g. maxTokens: 50) the entire budget goes to reasoning, finish_reason is length, and content is empty. Give reasoning prompts a few hundred tokens of headroom.

404 "model not found" for llama/qwen models​

Those models are retired, and so is gemma-4-31b. The catalog carries gpt-oss-120b only — verify the live roster with an authenticated GET https://api.cerebras.ai/v1/models.

See Also​

Need Help? Open a GitHub Discussion or issue.