Fireworks AI Free Credits (2026): Signup Credit Status

Short answer: Fireworks AI does not advertise free signup credits in 2026. It runs on prepaid credits: you add a payment method, buy credits, and every request is deducted from your balance. The one free offer on the official site is a promotion: free LoRA fine-tuning of Muse Glimmer 30B until November 30, 2026, with no credit card required. If you mainly want free tokens for inference, other providers are a better fit.

We list the signup credit as unverified: older third-party sites mention a $1 starter credit, but we found no mention of it on Fireworks' pricing page or docs on 2026-10-17.

What is actually free on Fireworks

Free Muse Glimmer 30B fine-tuning (until November 30, 2026)

Fireworks' pricing page says Muse Glimmer 30B SFT is free via the UI and Serverless Training API until 11/30, and that no credit card is required. The Muse Glimmer cookbook gives the details:

• What is covered: LoRA supervised fine-tuning (SFT) of Muse Glimmer 30B, run either as Managed SFT from the UI or through the Serverless Training API.
• Deadline: the starter script applies the promotion to LoRA SFT jobs started before December 1, 2026, 00:00 Pacific time, and refuses to launch after that unless you pass a flag to allow paid jobs.
• What is not covered: RL, DPO, full-parameter training, deployments, and inference outside the training run.
• Example data: eligible accounts get two preloaded example datasets: a 600-row prompt-router set and a 300-row receipts set, so you can run a first job without preparing data.
• Default setup: LoRA rank 8, base model accounts/fireworks/models/muse-glimmer-30b.

This is a genuinely useful offer if you have never fine-tuned a model. You can see the full workflow, from dataset to tuned adapter to a base-versus-tuned comparison, without paying for training. Just remember that serving the tuned model afterwards is billed like any other usage.

Access without a payment method

Fireworks' account quotas page shows that you can have an account without a payment method, but with tight limits: 10 requests per minute account-wide when there is no payment method or no credits, and 0 training GPUs until a payment method is on file. Because billing is prepaid, ordinary inference still needs a credit balance.

How to get started, step by step

1. Create an account at fireworks.ai.
2. For the free fine-tune only: open the fine-tuning section in the UI and start a Muse Glimmer 30B LoRA SFT job, or follow the cookbook to run it through the Serverless Training API. No card is needed.
3. For inference: add a valid payment method and billing profile on the Billing page. This puts you in Tier 1.
4. Buy prepaid credits. Optionally turn on Auto Reload to top up when your balance runs low.
5. Set a monthly spend limit so a bug cannot run up costs. Fireworks emails you at 80% of the limit by default.
6. Create an API key in the dashboard and call any serverless model with the OpenAI-compatible endpoint.

curl https://api.fireworks.ai/inference/v1/chat/completions \
  -H "Authorization: Bearer $FIREWORKS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "accounts/fireworks/models/gpt-oss-120b", "messages": [{"role": "user", "content": "Hello"}]}'

What it costs: serverless prices

Fireworks bills serverless inference per token, with separate prices for input, cached input and output. These are Standard mode prices per million tokens from the serverless pricing docs on 2026-10-17. Priority mode costs more; US-only variants of some models cost 50% more.

Models not listed individually are priced by size: $0.10 per million tokens under 4B parameters, $0.20 for 4B to 16B, and $0.90 above 16B (with separate rates for mixture-of-experts models). Embeddings cost $0.008 per million input tokens for models up to 150M parameters and $0.016 for 150M to 350M.

Worked example: a small test budget

Take 1,000 requests, each with 2,000 input tokens and 500 output tokens: 2 million input and 0.5 million output tokens in total, with no caching.

• Nemotron 3.5 Lightning: 2 x $0.05 + 0.5 x $0.20 = $0.20
• GPT OSS 120B: 2 x $0.15 + 0.5 x $0.60 = $0.60, or $0.30 through batch inference, which Fireworks bills at 50% of serverless prices
• DeepSeek V4.1 Flash: 2 x $0.30 + 0.5 x $1.20 = $1.20
• Kimi K3: 2 x $3.00 + 0.5 x $15.00 = $13.50

So $10 of credits covers roughly 16,000 such requests on GPT OSS 120B at Standard prices, or about 740 on Kimi K3. Prompt caching lowers these numbers further when requests share a long prefix.

Spending tiers and limits

• Tiers set your throughput ceiling. They control the maximum serverless tokens-per-minute your account can reach. Adding prepaid credits unlocks a higher tier within minutes.
• Request rate: 10 requests per minute with no payment method or no credits; up to 6,000 requests per minute account-wide with a payment method and active credits. The 6,000 cap is fixed and applies to all API usage.
• Monthly spend limit: on prepaid accounts this tracks the dollar value of your usage, even when credits pay for it. Adding credits does not raise or reset the limit.
• Auto Reload is separate from the spend limit. It buys credits when your balance is low; it does not change how much usage you can accrue in a month.

Fine-tuning prices after the promo

Once the Muse Glimmer promotion ends, or for any other model, fine-tuning is charged per million training tokens:

Muse Glimmer 30B falls in the 16.1B to 80B band, so a normally priced LoRA SFT run would be $3.00 per million training tokens. A run over 5 million training tokens would cost 5 x $3.00 = $15.00 at that rate, which is what the promotion saves you.

Free alternatives if you want free inference

Fireworks and Together AI are close competitors with similar open-model catalogs, and both are prepaid. Several models cost the same on both: GPT OSS 120B at $0.15 and $0.60, and GLM 5.3 at $1.40 and $4.40 per million tokens. For startups, larger credit programs such as NVIDIA Inception are covered in our guide on how to get free AI credits. For a full list of no-cost keys, see free AI API keys.

Who Fireworks is a good fit for

• Teams that want to fine-tune and serve open models in one place. The free Muse Glimmer run is a low-risk way to test that workflow.
• Apps that need high request rates. Up to 6,000 requests per minute account-wide is high for a self-serve plan.
• Cost-sensitive batch jobs. Batch inference at half price suits evaluation runs, data labeling and offline summarization.
• Not a fit if you need ongoing free inference for a hobby project. There is no free inference allowance to rely on.

History and common mistakes

• The $1 credit: older third-party listings say new Fireworks accounts got $1 in free credit. We could not find it on any current official page, so treat it as a past offer unless you see it in your own console.
• Billing model: Fireworks now describes itself as a prepaid credits system. Spend limits that older guides describe for postpaid accounts now apply only to legacy self-serve postpaid accounts.
• Mistake: assuming the promo covers serving. Free training does not include deployments or inference of the tuned model.
• Mistake: missing the deadline. Jobs must start before December 1, 2026, Pacific time. The year is not shown on the pricing page, only in the docs.
• Mistake: expecting credits to raise your spend cap. On prepaid accounts, adding credits does not raise or reset your monthly spend limit; change the limit itself.