LLM Strategy2026-10-2611 min read

Free LLM APIs for AI Trading Strategies

An AI trading strategy needs two things: data to read and a model to think with. The model is where most people expect a bill. It doesn't have to be one.

LLM Sessions and LLM Reactions on Sigrex run on your own API key from one of 21 supported providers. Several of them give API access for free: some with a daily request allowance, some with monthly credit, some with no published cap at all. That is enough to build, test and run a real AI strategy without paying for a single token.

This guide covers which providers are free right now, which of their models you can pick in Sigrex, how far each free tier stretches, and how to set things up so a free tier doesn't run dry halfway through the day. Everything here was checked on October 23–25, 2026. Free tiers change often, so check the provider's own page before you rely on a number.

First, count your requests

Free tiers are measured in requests per minute (RPM), requests per day (RPD) and sometimes tokens. So before you pick a provider, work out what your strategy will ask for.

An LLM Session runs on a schedule, and every run is at least one request:

Two things push the real number up.

Tool calls. With signal sending on, the model gets tools like get_candlestick_data, web_fetch and open_position. Every tool call is one more round trip to the model. A run that reads candles, checks the price and opens a position is three or four requests, not one.

Reactions. An LLM Reaction runs once for every event that reaches its Data Webhook. A busy feed can fire it hundreds of times a day. The reaction's frequency setting caps how often it can run.

The rule of thumb: your daily allowance divided by the requests per run is the number of runs you can afford. With 50 free requests a day and about 3 requests per run, that is 16 runs, so a 2-hour session.

The free providers at a glance

These are the providers in Sigrex's list with a free way in, best first for most strategies. Limits are what each provider published or reported when we checked.

"Reads charts?" means at least one model on that connection accepts images, which an LLM Session needs when chart input is on.

Provider by provider

NVIDIA (build.nvidia.com)

Join the free NVIDIA Developer Program on build.nvidia.com, create a key under API Keys (it starts with nvapi-), and the hosted catalog is free to call. NVIDIA doesn't publish a request cap for this tier. Developers on NVIDIA's forum consistently report about 40 requests per minute, and NVIDIA staff say the limit depends on the model and current traffic.

Models you can pick in Sigrex:

Text: nvidia/nemotron-3-ultra-550b-a55b, deepseek-ai/deepseek-v4.1-flash, openai/gpt-oss-120b, nvidia/nemotron-3.5-lightning-30b-a3b, poolside/laguna-xs-2.1
Reads charts: google/gemma-4-31b-it, z-ai/glm-5.3-flash, stepfun-ai/step-5-preview, nvidia/nemotron-3-nano-omni-30b-a3b-reasoning

NVIDIA describes this access as for prototyping, research, development and testing. It is the most generous free tier on this list, and the right place to build and forward test.

OpenRouter

Any model whose ID ends in :free costs nothing per token. Every account gets 20 requests per minute on free models. The daily cap is 50 until you have bought $10 of credits in total; after that it is 1,000 a day, and it stays at 1,000 even when your balance drops back to zero.

Free models in Sigrex's OpenRouter list that were being served on October 25:

google/gemma-4-31b-it:free and google/gemma-4-26b-a4b-it:free (both read charts)
nvidia/nemotron-3-ultra-550b-a55b:free
poolside/laguna-s-2.1:free and poolside/laguna-xs-2.1:free

OpenRouter adds and retires free models without much notice, so check the free model list before you build on one. Fifty requests a day carries roughly a 2-hour session with tools. With the one-time $10 purchase, 1,000 a day is enough for a 5-minute session.

Google Gemini

Create a key in Google AI Studio. The free tier is not a trial; it stays. On Google's pricing page these models in Sigrex's list are free of charge: gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite and gemini-3-flash-preview. gemini-4-argon and gemini-3.1-pro-preview are paid only.

Limits apply per Google Cloud project, and you see yours in AI Studio. Every Gemini model reads images, which makes this the easiest free option for chart sessions. On the free tier, Google may use your prompts and responses to improve its products.

Groq

Groq's free plan needs no card. For openai/gpt-oss-120b and openai/gpt-oss-20b it allows 30 requests a minute, 1,000 a day, 8,000 tokens a minute and 200,000 tokens a day. Groq is very fast, but the token limit is the one to watch: a prompt that pastes in a few hundred candles can pass 8,000 tokens on its own. Use it with lean prompts (see "Make a free tier last" below). Your exact limits are on the Limits page in the Groq console.

Nous Research

The Nous Portal has a Free plan: free models only, standard rate limits, no card. The Portal catalog marks several models at $0.00 per million tokens, and these are the ones in Sigrex's Nous list:

stepfun/step-3.7-flash (reads charts)
poolside/laguna-s-2.1 and poolside/laguna-xs-2.1
meituan/longcat-2.0
reflection/beam-1.0-501b-a23b
upstage/solar-mini4

Some of these also have a paid variant in the catalog, so check the price shown next to the model before you rely on it being free.

Z.ai

Z.ai's pricing page lists GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash as free for input and output. Sigrex offers GLM-4.6V-Flash, which reads images, so it is a free model for chart sessions. The other GLM models on Z.ai are paid, including GLM-5.3-Flash. If you want GLM-5.3-Flash for free, use it through NVIDIA instead.

SambaNova

Sign up at cloud.sambanova.ai and use it without adding a card. The free tier gives 20 requests a minute, 20 a day and 200,000 tokens a day per model. Twenty a day is tight for anything fast, but fine for a daily or 4-hour session, or a reaction to a rare event.

Models in Sigrex: DeepSeek-V3.1, DeepSeek-V3.2, Meta-Llama-3.3-70B-Instruct, gpt-oss-120b and gemma-4-31B-it (reads charts).

Mistral

Mistral's Experiment plan is free with a verified phone number, and you have to opt in to your API data being used for training. Mistral shows the exact limits in your console; trackers list them around 1 request per second per model. Sigrex offers mistral-small-latest, mistral-medium-latest and mistral-large-latest, and all three read images.

Cohere

Every Cohere account gets a trial key automatically. It allows 20 chat requests a minute and 1,000 API calls a month in total, and Cohere's terms say trial keys are for evaluation, not production or commercial use. A thousand calls is about 33 a day: enough to try command-a-03-2025 or command-a-plus-05-2026 on a 4-hour session.

Vercel AI Gateway

Every Vercel team gets $5 of AI Gateway credit every 30 days, starting from its first request, for as long as it never buys credits. It is credit rather than free models, so it reaches paid models too. Sigrex's Vercel list has deepseek/deepseek-v4-flash, zai/glm-5.3-flash, openai/gpt-5.6-luna and google/gemini-3.8-flash next to Claude and GPT flagships. On the cheap models $5 lasts a month of light use; on a flagship it can be gone within days.

Cerebras

Cerebras has no permanent free tier. New accounts get $5 of credit once they add a payment method, and it expires after 30 days. Trial limits for gpt-oss-120b are 5 requests a minute and 1 million tokens a day. Good for a short test of its speed, not for a long-running free setup.

Almost free: Meta's contributor tier

Meta's muse-spark-1.3-contributor costs $0.10 per million input tokens and $0.20 per million output tokens, in exchange for Meta training future models on your prompts and completions. Not free, but an hourly session with a typical prompt costs well under a dollar a month.

Connect a free provider in Sigrex

Create an API key on the provider's site (links in the table above).
In Sigrex, add a new LLM connection and choose the provider.
Pick the model. For a session with chart input, pick one that reads images.
Paste the key and click Create Connection. Sigrex sends one test request, and the connection is saved once the key works.
Select the connection when you create an LLM Session or an LLM Reaction.

Keys are encrypted and only decrypted when a strategy runs. You can switch the model on a connection whenever you like; for a different key, create a new connection. The full steps are in the docs: Connecting LLM Service.

Pair a free model with a fallback

Every LLM Session and LLM Reaction can have a second, fallback connection. If the main connection is unavailable, the run goes to the fallback instead. On free tiers this is worth doing every time: free models get busy at peak hours, and providers retire free models with little notice.

The best fallback is the same model at a different provider, so your strategy behaves the same whichever one answers:

If you can't find the same model on two free providers, a similar one works too: Gemini gemini-3.8-flash with OpenRouter google/gemma-4-31b-it:free as the fallback keeps chart reading free on both sides.

Make a free tier last

Match the interval to the allowance. Use the table at the top. A 5-minute session on a 50-a-day tier runs dry before breakfast.

Put the data in the prompt instead of letting the model fetch it. {{get:url}} fetches data before the model runs, so it costs no extra request. The web_fetch tool costs a full round trip every time the model uses it. The Sigrex changelog calls {{get:}} the most cost-efficient option for requests you know in advance.

Shrink what you paste in. Wrap JSON in {{toon: ... }} to convert it to the compact TOON format, which uses fewer tokens. Ask APIs for 24 candles, not 500 (limit=24), and for one symbol, not the whole market. This matters most on Groq (8,000 tokens a minute) and SambaNova (200,000 tokens a day).

Keep the journal short. Strategy storage is added to the model's instructions on every run, so a journal that only ever grows makes every request bigger. Tell the model to keep the last 10 to 20 entries.

Let code do the watching. A Code Strategy loops continuously and uses no LLM at all. Let it watch the price and post to a Data Webhook only when something worth a look happens. An LLM Reaction on that webhook then thinks only when there is something to think about.

Start with signals off. With signal sending off the model gets no tools, so each run is a single request. Read its analysis for a few days, then turn signals on.

Three free setups to copy

1. An hourly chart reader on Gemini

Connection: Google Gemini, gemini-3.8-flash. Fallback: OpenRouter google/gemma-4-31b-it:free.
Session: every 1 hour, chart input on with BINANCE:BTCUSDT, 1h candles, RSI and EMA. Send email on, signal sending off.
Cost: 24 requests a day, $0.

{{comment: Chart input ON: BINANCE:BTCUSDT, 1h candles, RSI + EMA}}
You are reviewing the BTCUSDT 1h chart for a swing trader.
Time: {{current_time}}
Last price: {{get:https://api.binance.com/api/v3/ticker/price?symbol=BTCUSDT}}

From the chart only:
1. Trend: is price above or below the EMA, and is the EMA rising or falling?
2. Momentum: the RSI value and zone (below 30, 30 to 70, above 70).
3. The nearest support and resistance you can see.

Answer in five lines or fewer. End with one word: BULLISH, BEARISH or NEUTRAL.

2. A 15-minute signal strategy on NVIDIA

Connection: NVIDIA, nvidia/nemotron-3-ultra-550b-a55b. Fallback: OpenRouter nvidia/nemotron-3-ultra-550b-a55b:free.
Session: every 15 minutes, signal sending on, signal symbol BTCUSDT. Point it at a Bot Webhook with no bot attached for the first week, so every signal is logged and nothing trades (how to forward test).
Cost: 96 runs a day, roughly 300 requests with tool calls. Well inside NVIDIA's rate; the OpenRouter fallback only steps in when NVIDIA is unavailable.

{{val:symbol=BTCUSDT}}
You trade {{symbol}} perpetual futures on the 15m timeframe. Take only clear setups; staying flat is a valid answer.

Last 24 candles (15m, oldest first; open time, open, high, low, close, volume, ...):
{{toon: {{get:https://api.binance.com/api/v3/klines?symbol={{symbol}}&interval=15m&limit=24}} }}

Funding and mark price:
{{toon: {{get:https://fapi.binance.com/fapi/v1/premiumIndex?symbol={{symbol}}}} }}

Fear & Greed:
{{toon: {{get:https://api.alternative.me/fng/?limit=1}} }}

Rules:
- Open LONG only if the last 3 closes are above the 20-candle average close and lastFundingRate is below 0.0003.
- Open SHORT only if the last 3 closes are below the 20-candle average close and lastFundingRate is above -0.0003.
- Close an open position at +1.2% or -0.8% from the entry price.
- Otherwise do nothing.

Keep storage as one object: {"entry": number or null, "side": "LONG", "SHORT" or null, "notes": the last 10 one-line notes}.

The rules are an example to show the format, not a strategy we recommend. Sigrex adds its own system prompt when signals are on, which tells the model there can be only one open position at a time.

3. A free daily market brief on SambaNova

Connection: SambaNova, DeepSeek-V3.1. Fallback: Groq openai/gpt-oss-120b.
Session: daily, send email on, signal sending off.
Cost: one request a day.

Write my morning crypto brief in plain language, under 200 words.

Fear & Greed, last 7 days:
{{toon: {{get:https://api.alternative.me/fng/?limit=7}} }}

BTC and ETH, last 24 hours:
{{toon: {{get:https://api.binance.com/api/v3/ticker/24hr?symbols=["BTCUSDT","ETHUSDT"]&type=MINI}} }}

BTC funding, last 3 periods:
{{toon: {{get:https://api.bybit.com/v5/market/funding/history?category=linear&symbol=BTCUSDT&limit=3}} }}

Altcoin Season Index (0-100):
{{get:https://pro-api.coinmarketcap.com/public-api/v1/altcoin-season-index/latest}}

Cover the overall mood and how it changed this week, the BTC and ETH moves, whether funding looks crowded, and whether altcoins are leading. No predictions, no advice.

Every URL in these prompts is free and needs no key. Our next post lists many more of them, from exchange order books to prediction-market odds.

The fine print

Free tiers change. Every number here was read between October 23 and 25, 2026. Providers raise, cut and rename free tiers, sometimes overnight.
Some free tiers pay with your data. Gemini's free tier, Mistral's Experiment plan and Meta's contributor tier may use your prompts to improve or train their models. Keep anything private out of prompts on those tiers.
Some are meant for testing. NVIDIA's free endpoints are for prototyping and development, and Cohere's trial keys exclude production and commercial use. Read the terms before real money depends on one.
Free models are slower and sometimes busy. That is what the fallback connection is for.
A free model is still a model. It can misread a chart or talk itself into a trade. Forward test every AI strategy on a webhook with no bot attached before it touches your account.

Start building

The Sigrex free plan includes 2 LLM Sessions and 2 LLM Reactions, with no card. Connect one of the free providers above and you can run an AI strategy for 0€ on both sides.

Create a free account and connect your first free model.