> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.jambonz.org/guides/features/bring-your-own-llm/hugging-face/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.jambonz.org/_mcp/server.

# HuggingFace

> **Note**
>
> This vendor specifically integrates with **HuggingFace Inference Providers** — HF's multi-provider broker at `https://router.huggingface.co/v1`. It is **not** for HuggingFace's older shared Inference API (a different wire protocol) or dedicated Inference Endpoints (those are per-customer URLs that work with `vendor: openai` and a `baseURL` override).

HuggingFace Inference Providers is a multi-provider broker. One HF token at `https://router.huggingface.co/v1` routes requests to whichever inference shop hosts the requested model — Groq, Together, Fireworks, Cerebras, Nebius, Hyperbolic, SambaNova, etc. The wire is OpenAI-compatible.

**Why pick this over a direct vendor?** A customer who wants Llama on Groq, Mistral on Together, and Qwen on Fireworks needs three separate accounts and three jambonz credentials when going direct. With HF Providers, *one* HF token covers all of them, and you switch backends just by changing the model id.

The tradeoff: you don't get the absolute cheapest per-token pricing — HF takes a small markup on top of provider rates. For production scale, going direct to your chosen provider may save money. For prototyping and multi-vendor testing, HF wins on convenience.

## Get credentials

1. Sign in at [https://huggingface.co](https://huggingface.co).
2. Open **Settings → Access Tokens** (direct link: [https://huggingface.co/settings/tokens](https://huggingface.co/settings/tokens)).
3. Click **+ Create new token**. Name it (e.g. `jambonz`). Pick token type **Read** — that's enough for Inference Providers.
4. Copy the `hf_...` string.

> **Warning**
>
> Inference Providers requires an **inference credit balance**. HF gives a small free monthly allowance; for production use, attach a payment method at [https://huggingface.co/billing](https://huggingface.co/billing). Without credits, requests return `402 Payment Required`.

## Configure in jambonz

In the portal: **Account → LLM Services → + Add LLM Service → HuggingFace Inference Providers**.

**`HuggingFace Token`** `string` — required

The `hf_...` token from your HF account settings.

---

**`Base URL`** `string`

Defaults to `https://router.huggingface.co/v1`. Override only for proxies.

---

Click **Test** to verify.

## Use in an agent verb

```js
session.agent({
  llm: {
    vendor: 'huggingface',
    model: 'meta-llama/Llama-3.3-70B-Instruct',
    llmOptions: {
      systemPrompt: 'You are a helpful voice assistant.',
    },
  },
  stt: { vendor: 'deepgram', language: 'en-US' },
  tts: { vendor: 'cartesia', voice: 'sonic-english' },
  turnDetection: 'krisp',
  bargeIn: { enable: true },
  actionHook: '/agent-complete',
}).send();
```

## Model field syntax

The `model` value is a canonical HF model id (`<org>/<model>`), optionally with a routing-hint suffix:

| Form                                             | Behavior                                               |
| ------------------------------------------------ | ------------------------------------------------------ |
| `meta-llama/Llama-3.3-70B-Instruct`              | HF picks a backend (often Groq for fast Llama models). |
| `meta-llama/Llama-3.3-70B-Instruct:fastest`      | HF picks the lowest-latency backend at request time.   |
| `meta-llama/Llama-3.3-70B-Instruct:fireworks-ai` | Pin Fireworks.                                         |
| `meta-llama/Llama-3.3-70B-Instruct:cerebras`     | Pin Cerebras.                                          |
| `meta-llama/Llama-3.3-70B-Instruct:groq`         | Pin Groq.                                              |

The provider chosen is reported in the `x-inference-provider` response header. Useful for diagnosing latency anomalies — if you expected Cerebras and got Groq, you know the routing kicked in differently.

## Available Models

Browse [HuggingFace's full Inference Providers model list](https://huggingface.co/inference/models) (filterable by provider) or read the [Inference Providers documentation](https://huggingface.co/docs/inference-providers/en/index) for the routing model. Common picks for voice agents:

| Model                                | Notes                              |
| ------------------------------------ | ---------------------------------- |
| `meta-llama/Llama-3.3-70B-Instruct`  | General-purpose, often Groq-routed |
| `meta-llama/Llama-3.1-8B-Instruct`   | Smaller, cheaper                   |
| `Qwen/Qwen2.5-72B-Instruct`          | Strong reasoning                   |
| `mistralai/Mistral-7B-Instruct-v0.3` | Mistral via partner                |
| `deepseek-ai/DeepSeek-V3`            | DeepSeek through HF                |

## Quirks & errors

> **Warning**
>
> **`402 Payment Required`** — your HF account has no inference credits. Either claim the free monthly allowance or attach a payment method at [huggingface.co/billing](https://huggingface.co/billing).

> **Warning**
>
> **`410 Gone — model is deprecated and no longer supported by provider 'X'`** — provider pinning works only for currently-hosted pairs. If a specific provider drops support for a model, requests pinned to that provider fail. Drop the suffix (let HF auto-route) or pick a different provider.

> **Note**
>
> **Provider routing is dynamic.** A request without a suffix today might route to Groq; tomorrow it could route to Cerebras if HF's broker decides differently. If you need deterministic routing for benchmarking or compliance, pin a specific provider with the `:provider-name` suffix.

> **Note**
>
> **Read scope on the token is enough.** Higher scopes (Write, Admin) work too but are over-privileged for inference-only use.

> One HuggingFace token, many backends — Groq, Together, Fireworks, Cerebras, Nebius all behind a single broker endpoint.