Together AI Free code examples in Python and JavaScript

THE PACKAGE IS together-ai, with the suffix. The bare `together` name on npm is an unrelated project.

npm install together-ai

export TOGETHER_API_KEY="your-key"

The response is OpenAI-shaped. Streaming is server-sent events — pass stream: true and iterate.

import Together from "together-ai";

const client = new Together({ apiKey: process.env.TOGETHER_API_KEY });

const completion = await client.chat.completions.create({
  model: "meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
  messages: [{ role: "user", content: "Explain a database index." }],
});

console.log(completion.choices[0].message.content);

JSONL, one conversation per line, matching what you will send at inference. A few hundred representative examples beat a large scraped set — and the mismatch between training shape and request shape is the usual reason a fine-tune underperforms.

import { writeFile } from "node:fs/promises";

const examples = [
  {
    messages: [
      { role: "system", content: "You triage pump faults." },
      { role: "user", content: "Pump stalled, code E-4471." },
      { role: "assistant", content: "SEVERITY: high\nCAUSE: impeller jam" },
    ],
  },
];

await writeFile(
  "train.jsonl",
  examples.map((example) => JSON.stringify(example)).join("\n") + "\n",
);

const uploaded = await client.files.upload({
  file: "train.jsonl",
  purpose: "fine-tune",
});

LoRA first, always. And BEFORE running a job at all: if a better prompt or a retrieval step gets you there, that is cheaper, reversible and does not go stale.

const job = await client.fineTune.create({
  training_file: uploaded.id,
  model: "meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
  lora: true,
  n_epochs: 3,
  suffix: "pump-triage",
});

// Worth fine-tuning for:
//   a fixed output shape the model keeps drifting from
//   a house style or tone that a prompt cannot hold across turns
//   domain vocabulary the base model mangles
//
// NOT worth fine-tuning for:
//   facts that change          -> retrieval
//   a one-off instruction      -> the prompt
//   "it is not smart enough"   -> a bigger base model
console.log(job.id, job.status);

The finished job's output name is a model id like any other. Keep the base model's call alongside it for a while — a fine-tune that is better on your examples and worse in general is a common and quiet outcome.

let status;
do {
  await new Promise((resolve) => setTimeout(resolve, 30_000));
  status = await client.fineTune.retrieve(job.id);
  console.log(status.status);
} while (!["completed", "error", "cancelled"].includes(status.status));

if (status.status !== "completed") {
  throw new Error(`fine-tune ${status.status}`);
}

// Run both and compare on held-out examples before switching.
for (const model of [
  "meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
  status.output_name,
]) {
  const completion = await client.chat.completions.create({
    model,
    messages: [{ role: "user", content: "Pump stalled, code E-4471." }],
  });
  console.log(model, "->", completion.choices[0].message.content);
}