Skip to main content

Featherless AI Provider Guide

Featherless AI 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/featherless-ai.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. Every field below comes from Featherless's own published docs and an unauthenticated GET /v1/models call — no API key was used to write this entry, so no request-response wire behavior (streaming, tool calling, structured output) has actually been exercised end to end yet. See Verification status for exactly what was and was not checked.

Key Facts​

  • Provider id: featherless-ai
  • Protocol: OpenAI-compatible (/chat/completions), per Featherless's own docs: "Featherless's API is designed to be API compatible with OpenAI."
  • Base URL: https://api.featherless.ai/v1
  • Default model: unsloth/Llama-3.3-70B-Instruct
  • Models in catalog: 12 (out of a live roster of 22,060 models — see Models for why these 12)
  • Streaming: not documented anywhere in Featherless's docs — declared false until confirmed live
  • Tool calling: model-dependent. Featherless's Tool Calling guide natively supports only the moonshotai/Kimi-K2-Instruct and Qwen 3 model families; other models need a prompting workaround the guide documents separately
  • Structured output: not documented (no response_format/JSON-mode mention anywhere in the docs checked) — declared false
  • Embeddings: Featherless does document a separate POST /v1/embeddings endpoint for embedding-output models, but that is a distinct model family from the chat models in this catalog entry and is not wired through this provider — declared false
  • Billing: no free tier. Plans start at $25/mo
  • Key format: none declared by the vendor
  • Gated models: some roster ids are Hugging-Face-gated and return HTTP 403 until unlocked manually in the account UI (see Troubleshooting)

Quick Start​

1. Get an API key​

  1. Visit: https://featherless.ai/register and create an account (or https://featherless.ai/login if you already have one)
  2. Subscribe to a plan — Chat ($25/mo, unlimited tokens, 32K context) or Developer ($50/mo in credits, billed per token, up to 256K context); Featherless has no free or keyless tier
  3. Generate an API key from the API keys section of your account
  4. Set FEATHERLESS_AI_API_KEY in your .env file

2. Configure​

export FEATHERLESS_AI_API_KEY=your-api-key
export FEATHERLESS_AI_MODEL=unsloth/Llama-3.3-70B-Instruct   # optional — overrides the default model
export FEATHERLESS_AI_BASE_URL=https://api.featherless.ai/v1 # optional — self-hosted or proxy

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: "featherless-ai",
  model: "unsloth/Llama-3.3-70B-Instruct",
});

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

Per-request credentials work as they do for every provider:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "featherless-ai",
  credentials: {
    "featherless-ai": { apiKey: process.env.FEATHERLESS_AI_API_KEY },
  },
});

Models​

Featherless's live roster is enormous — 22,060 models as of 2026-09-28, spanning mainline instruction-tuned checkpoints alongside a very large number of community roleplay, uncensored and "abliterated" (safety-refusal-removed) fine-tunes; Featherless's own site markets these as first-class catalog filters ("LLMs for Roleplay", "Uncensored Models", "Abliterated Models"). This catalog entry lists 12 models — a bounded, verifiable slice, not the whole roster — chosen to include the default, its closest official counterpart, and a representative spread of the architectures Featherless serves (Llama-3.3-70B family, QRWKV, RWKV5).

ModelContextVision$/M in · outNotes
unsloth/Llama-3.3-70B-Instruct ⭐32Kno$2.60 / $3.00Recommended — an ungated community re-upload of Meta's Llama 3.3 70B Instruct. Chosen as default because the official meta-llama id is gated (see below). Roster reports tool_use: true.
meta-llama/Llama-3.3-70B-Instruct32Kno$0.65 / $0.75Meta's official weights. Hugging-Face-gated on Featherless (HTTP 403 until unlocked in the account UI) — kept as a fallback, not the default.
Sao10K/L3.3-70B-Euryale-v2.332Kno$0.65 / $0.75Roleplay/long-form-writing Llama-3.3-70B fine-tune. Not gated; roster-listed only.
EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.032Kno$2.60 / $3.00Creative-writing/roleplay Llama-3.3-70B fine-tune. Not gated; roster-listed only.
EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.132Kno$2.60 / $3.00Later revision of the EVA-LLaMA line above. Not gated; roster-listed only.
KaraKaraWitch/Llama-MiraiFanfare-3.3-70B32Kno$2.60 / $3.00Community Llama-3.3-70B fine-tune. Not gated; roster-listed only.
huihui-ai/Llama-3.3-70B-Instruct-abliterated32Kno$0.65 / $0.75Safety-refusal-ablated Llama 3.3 70B Instruct. Not gated; roster-listed only.
recursal/QRWKV6-32B-Instruct-Preview-v0.132Kno$0.27 / $0.65Non-transformer RWKV-hybrid architecture; one of the few model classes Featherless's docs call out at the 32K context tier. Not gated; roster-listed only.
RWKV/v6-Finch-14B-HF16Kno$0.11 / $0.28RWKV v6 "Finch" 14B, non-transformer architecture. Not gated; roster-listed only.
RWKV/EagleX-7B-Chat-V0.5-pth16Kno$0.10 / $0.20Early RWKV5-architecture chat checkpoint. Not gated; roster-listed only.
recursal/EagleX_1-7T_Chat16Kno$0.10 / $0.20RWKV5-architecture, chat-tuned. Not gated; roster-listed only.
recursal/EagleX_1-7T16Kno$0.10 / $0.20RWKV5-architecture base checkpoint (non-chat-tuned). Not gated; roster-listed only.

Context window and pricing come directly from the live, unauthenticated GET /v1/models response (context_length, pricing.input/pricing.output, already expressed as $ per million tokens) fetched 2026-09-28 — not from invented numbers. Featherless does not document a per-request output-token ceiling anywhere in its docs (max_tokens is described only as "Maximum number of tokens generated per output sequence", with no number given), so defaultMaxOutputTokens (4096) follows the same conservative, undocumented-cap convention already used by several other catalog entries (Fireworks, Groq, Mistral, Novita, Perplexity, Together AI, xAI) rather than a Featherless-specific figure.

Fallback order when the default is unavailable: meta-llama/Llama-3.3-70B-Instruct → Sao10K/L3.3-70B-Euryale-v2.3 → EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.0 → EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.1 → KaraKaraWitch/Llama-MiraiFanfare-3.3-70B → huihui-ai/Llama-3.3-70B-Instruct-abliterated → recursal/QRWKV6-32B-Instruct-Preview-v0.1 → RWKV/v6-Finch-14B-HF → RWKV/EagleX-7B-Chat-V0.5-pth → recursal/EagleX_1-7T_Chat → recursal/EagleX_1-7T.

Verification status​

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This entry was built entirely credential-free — no signup, no API key, no POST request to Featherless was made — so it records what a public, unauthenticated pass can prove and nothing more:

ProbeResult
Rosterunauthenticated GET /v1/models, HTTP 200, 22,060 models, 2026-09-28
Billingunauthenticated GET https://featherless.ai/pricing, HTTP 200, 2026-09-28 — no free tier found anywhere on the page; plans start at $25/mo
Auth rejectionnot probed — would require a POST request, which credential-free onboarding does not permit
Live capability sweepnot run (evidence.liveMatrix is null) — chat, streaming, tool calling and structured output are all unverified against a real request
Docs pages readOverview, API overview and common options, Completions (/v1/chat/completions and /v1/completions), Quickstart guide, API examples and snippets, Tool Calling, Embeddings, Vision, Chat template kwargs, Models & Model Compatibility, Error codes, Pricing — none of these documents response_format or stream:true for chat/text completions

Because of this, capabilities.streaming, capabilities.structuredOutput, and capabilities.structuredOutputWithTools are all conservatively false, and capabilities.tools is "model-dependent" (Featherless's own Tool Calling guide frames it that way, naming only moonshotai/Kimi-K2-Instruct and the Qwen 3 family as natively supported). A live-verification pass with a real API key — npx tsx test/continuous-test-suite-provider-matrix.ts --provider=featherless-ai — is required before any of these flags should be promoted, per Tier-2 onboarding: Live verification.

Troubleshooting​

SymptomCauseFix
Invalid Featherless API keyFEATHERLESS_AI_API_KEY unset or wrongCheck the key in your Featherless account at https://featherless.ai/register
HTTP 403, "model is gated"The requested model (e.g. meta-llama/Llama-3.3-70B-Instruct) is Hugging-Face-gated on FeatherlessOpen the model's page on featherless.ai while signed in and click "Unlock Model" to accept its license, or use the ungated unsloth/Llama-3.3-70B-Instruct default instead
Model not foundThe roster changed since 2026-09-28, or the id was mistypedPick a current id from the unauthenticated GET /v1/models roster or https://featherless.ai/models
HTTP 400, "This Model is Cold"Featherless documents cold/loading/warm model states; a cold model needs to loadRetry — warm-up can take from a few minutes up to an hour for very large models, per Featherless's error-codes page
HTTP 503, "Service Temporarily Unavailable"Featherless documents this as insufficient capacity, not downtimeRetry the same request; Featherless's own guidance is to retry up to three times
Unexpected billingNo free or keyless tier existsConfirm plan choice at https://featherless.ai/#pricing before heavy use — Chat is a flat $25/mo, Developer is metered credits

See also​