Build AI apps on free tiers. 10 AI/ML infrastructure categories — LLM APIs, vector databases, experiment tracking, observability, and more. Exact limits, scaling guidance.

AgentDeals › Alternatives › Free AI/ML Stack

The Complete Free AI/ML Stack

Published 2026-03-25 · Not yet reviewed

Everything you need to build, train, and deploy AI applications — without spending a dollar. This guide recommends the best free tier for each layer of an AI/ML development stack — 10 categories from LLM APIs to speech AI — with exact limits for each. Figures in the tables below come from our records for 1,498 developer tools.

Designed for solo AI developers, indie hackers, and startup teams prototyping AI features. Each recommendation includes alternatives, a "when you'll outgrow it" guide, and stability notes based on our tracking of 513 real pricing changes. The limits on this page were read from vendor pricing pages between 2026-07-10 and 2026-11-13.

$0/month
Total estimated monthly cost for prototyping and experimentation
$0 covers the 4 of 10 picks whose free tier our own badge still reads as active. It does not cover Groq (at risk), GitHub Copilot (at risk), Hugging Face (at risk), Roboflow (at risk), Portkey (unrated — change not reconciled), Deepgram (credits only).

Stack Overview

Category Recommended Key Limit Stability
🧠 LLM API Access Groq Fast LLM inference on Groq's LPU hardware. Free plan: gpt-oss-20b, gpt-oss-120b and Qwen3.8 27B, each at 30… at risk
💻 AI Coding Assistant GitHub Copilot AI coding assistant by GitHub. Copilot Free is $0 with no credit card: 2,000 code completions a month, a… at risk
📐 Vector Database Pinecone Vector database. The Starter plan is free: up to 2 GB storage, 2M write units and 1M read units a month, 1 GB… active
📊 ML Experiment Tracking Weights & Biases ML experiment tracking, model registry, and LLM app tracing and evaluation (Weave). W&B's hosted service is… active
🚀 Model Hosting & Inference Hugging Face ML model hub. Since 2026-10-07, free accounts get no included Inference Providers credits: free users buy… at risk
⚡ Compute & Training Kaggle Google's data science platform with free notebooks, datasets and competitions. Notebooks get a weekly GPU… active
🏷️ Data Labeling & Annotation Roboflow Computer vision platform for labeling, training and deploying models. On September 18, 2026 Roboflow… not confirmed · page states a free price · 2026-11-02 at risk
🔍 AI Observability & Evaluation Langfuse Open-source LLM engineering platform for tracing, evaluating and debugging AI applications. The Langfuse… active
🔀 AI Gateway & Routing Portkey Control panel for Gen AI apps featuring an observability suite & an AI gateway. Send & log up to 10,000… not confirmed · read not reconciled · 2026-10-24 unrated — change not reconciled
🎙️ Speech & Audio AI Deepgram Speech-to-text and text-to-speech API — $200 free credits on signup (no credit card required), ~43K minutes… credits only

🧠 LLM API Access

Recommended Groq Free at risk

Ultra-fast inference on LPU hardware — 30 RPM, 1,000 requests and 200K tokens a day per model, free. Serves gpt-oss-120b, gpt-oss-20b and Qwen3.8 27B. Best balance of speed, limits, and model quality for prototyping.

Fast LLM inference on Groq's LPU hardware. Free plan: gpt-oss-20b, gpt-oss-120b and Qwen3.8 27B, each at 30 requests a minute, 1,000 requests and 200,000 tokens a day, plus Whisper speech-to-text at 2,000 requests a…

Full profile Alternatives console.groq.com/docs/rate-limits ↗

Also consider:

Google Gemini API Free (Reduced) active Mistral AI Pay-as-you-go unrated — not a free offer Cohere Free active OpenRouter Free active

No longer a free-tier pick: Cerebras — free tier removed.

When you'll outgrow it: When you exceed 30 RPM or 200K tokens a day on a model. At that point, OpenRouter (25+ free models) extends the free runway. Production apps typically need paid tiers for reliability SLAs.
Full comparison guide →

💻 AI Coding Assistant

Recommended GitHub Copilot Free at risk

2,000 completions/month free. VS Code native with deep GitHub integration. The simplest on-ramp to AI-assisted development.

AI coding assistant by GitHub. Copilot Free is $0 with no credit card: 2,000 code completions a month, a limited monthly allowance of GitHub AI Credits for chat and agent features, and models through auto model…

Full profile Alternatives github.com/features/copilot ↗

Also consider:

Cursor Free at risk Amazon Q Developer Free unrated — change refused Cline Open Source unrated — page states no price Aider Open Source unrated — page states no price
When you'll outgrow it: When you exceed 2,000 completions/month or need unlimited chat. Cline and Aider are free with your own API keys — no completion limits, just API costs.
Full comparison guide →

📐 Vector Database

Recommended Pinecone Free active

2 GB storage with 5 indexes and 2M write units/month on the Starter plan. Serverless architecture means zero ops. The most popular vector DB with excellent SDK support.

Vector database. The Starter plan is free: up to 2 GB storage, 2M write units and 1M read units a month, 1 GB of egress a month (reads that return data stop at the cap until the next billing period), 5 indexes, 1…

Full profile Alternatives pinecone.io/pricing ↗

Also consider:

Qdrant Free active
When you'll outgrow it: When you exceed 2 GB storage or need more than 5 indexes. Qdrant Cloud's free cluster (1 GB RAM, 4 GB disk) is an alternative; Qdrant suspends it after a week without use.
Full comparison guide →

📊 ML Experiment Tracking

Recommended Weights & Biases Free active

Free cloud plan for personal development: up to 5 model seats, 5 GB/month of storage and 1 GB/month of Weave ingestion. Industry-standard experiment tracking.

ML experiment tracking, model registry, and LLM app tracing and evaluation (Weave). W&B's hosted service is now sold as CoreWeave Forge. Its Free plan is $0/mo and designed for personal development, with up to 5 model…

Full profile Alternatives coreweave.com/forge-pricing ↗

Also consider:

Comet ML Free active
When you'll outgrow it: When you need more than 5 GB/month of storage or more than 5 model seats. W&B Pro starts at $60/month with a 30-day free trial.
Full comparison guide →

🚀 Model Hosting & Inference

Recommended Hugging Face Free at risk

The largest open-source model hub, free with 100GB of private storage — deploy models, share datasets, and collaborate on ML projects. Its Inference Providers API serves 200+ models; since 2026-10-07, free accounts buy credits to call it, or upgrade to PRO ($9/month, $2.00 a month of credits).

ML model hub. Since 2026-10-07, free accounts get no included Inference Providers credits: free users buy credits or upgrade to PRO to call Inference Providers, which serves 200+ models. PRO ($9/month) gets $2.00 a…

Full profile Alternatives huggingface.co/docs/inference-providers/pricing ↗

Also consider:

Replicate Trial trial only
When you'll outgrow it: When you need dedicated endpoints or higher throughput: production apps can use Inference Endpoints (dedicated, from $0.033/hour).
Full comparison guide →

⚡ Compute & Training

Recommended Kaggle Free active

A weekly quota of 30 GPU hours (one P100 or two T4s) and up to 20 TPU hours, at no charge. Integrated datasets and community notebooks.

Google's data science platform with free notebooks, datasets and competitions. Notebooks get a weekly GPU quota of 30 hours, sometimes more depending on demand (one NVIDIA Tesla P100 or two Tesla T4s), and a TPU v3-8…

Full profile Alternatives www.kaggle.com/docs/efficient-gpu-usage ↗

Also consider:

Google Colab Free active
When you'll outgrow it: When you need longer sessions (Kaggle caps them at 12 hours on CPU or GPU and 9 on TPU), more VRAM (T4 = 16 GB), or persistent storage. Google Colab offers T4 free but with session time limits.
Full comparison guide →

🏷️ Data Labeling & Annotation

Recommended Roboflow Free at risk

Free Core plan: 10 credits that refresh every month, private projects and models, and model weight download. Best for computer vision projects with built-in model training and deployment.

Computer vision platform for labeling, training and deploying models. On September 18, 2026 Roboflow announced a free tier for its Core plan: 10 credits that refresh every month, private projects and models, and model… not confirmed · page states a free price · 2026-11-02

Full profile Alternatives docs.roboflow.com/platform/billing-and-plans/plans ↗

Also consider:

Labelbox Free active
When you'll outgrow it: When you need more than the 10 monthly credits. Labelbox offers 500 LBUs/month.
Full comparison guide →

🔍 AI Observability & Evaluation

Recommended Langfuse Free active

50K units/month free on Cloud Hobby (traces, observations and scores each count), all platform features with limits, 30 days of data access, open-source. Trace LLM calls, evaluate outputs, manage prompts, and debug chains — the emerging standard for LLM ops.

Open-source LLM engineering platform for tracing, evaluating and debugging AI applications. The Langfuse Cloud Hobby plan is free with no credit card: 50k units a month (each trace, observation and score counts as one…

Full profile Alternatives langfuse.com/pricing ↗

Also consider:

Langtrace Free unrated — page unreachable Braintrust Free active LangWatch Free unrated — change not reconciled Portkey Free unrated — change not reconciled
When you'll outgrow it: When you exceed 50K units/month; Langfuse Core is $29/month with 100K units. Langtrace (50K traces/month) and LangWatch (1K traces/month) are alternatives. For production, self-host Langfuse (open-source) for unlimited traces.
Full comparison guide →

🔀 AI Gateway & Routing

Recommended Portkey Free unrated — change not reconciled

10,000 requests/month free. AI gateway with load balancing, fallbacks, semantic caching, and 200+ LLM providers. Route between models for cost optimization.

Control panel for Gen AI apps featuring an observability suite & an AI gateway. Send & log up to 10,000 requests for free every month. not confirmed · read not reconciled · 2026-10-24

Full profile Alternatives portkey.ai ↗

Also consider:

Keywords AI Free unrated — page unreadable Composio Free active
When you'll outgrow it: When you exceed 10K requests/month. Keywords AI also offers 10K requests/month free. For self-hosted, LiteLLM (open-source) has no limits.
Full comparison guide →

🎙️ Speech & Audio AI

Recommended Deepgram Free Credits credits only

$200 free credits on signup (~43K minutes transcription). Industry-leading speech-to-text accuracy with real-time streaming, multiple languages, and speaker diarization.

Speech-to-text and text-to-speech API — $200 free credits on signup (no credit card required), ~43K minutes transcription with Nova model, credits never expire. Access to all model endpoints

  • Deepgram's docs say only data included through its Model Improvement Partnership Program is stored and used for future model training. Adding mip_opt_out=true to a request excludes it, and opted-out request data is kept only while the request is processed. (From developers.deepgram.com/docs/the-deepgram-model-improvement-partnership-program, read 2026-10-23.)
  • Nova-3 speech-to-text allows up to 50 concurrent pre-recorded requests and up to 150 concurrent streaming requests. (From developers.deepgram.com/reference/api-rate-limits, read 2026-10-23.)
  • A transcription request without enough credits fails with HTTP 402 Payment Required, with the message "Project does not have enough credits for an ASR request and does not have an overage agreement." (From developers.deepgram.com/docs/errors, read 2026-10-23.)
  • Rate limits apply per project, not per account or API key. Creating additional projects under the same account does not grant additional concurrency. Secondary projects on a self-serve account are limited to a single concurrent stream. Using additional projects or accounts to work around the limits is detected and violates Deepgram's Terms of Service. (From developers.deepgram.com/reference/api-rate-limits, read 2026-11-07.)
  • Deepgram's terms, section 2.4, forbid use of the Services or any Output "for competitive purposes, including model training, benchmarking and other competitive analysis, or developing competing models, products or services". (From deepgram.com/terms, read 2026-11-07.)
  • Deepgram's terms, section 2.4, prohibit resale or redistribution of the Services on a stand-alone basis. Integration into the user's own Application is allowed when the Application contributes material independent functionality beyond Deepgram's Services. (From deepgram.com/terms, read 2026-11-07.)
Full profile Alternatives deepgram.com/pricing ↗

Also consider:

AssemblyAI Free Credits credits only
When you'll outgrow it: When you exhaust $200 credits. AssemblyAI offers $50 free credits (~185 hrs). For ongoing free usage, Whisper (open-source) can run locally via Hugging Face.
Full comparison guide →

Stability Notes

Recent pricing changes affecting vendors in this stack. Based on our tracking of 513 deal changes across 1,498+ developer tools.

pricing restructured Cursor: From October 12, 2026, new subscribers buy Cursor as part of The X Subscription, one plan for Grok, Cursor and X: Free $0, Plus $8, Premi... Source ↗
limits reduced Hugging Face: Hugging Face stopped giving free accounts included monthly Inference Providers credits; free users now have to buy credits or upgrade to ... Source ↗
limits reduced Groq: Groq cut the free limits of gpt-oss-safeguard-20b from 30 to 3 requests a minute and from 8,000 to 2,000 tokens a minute; 1,000 requests ... Source ↗
restriction Google Gemini API: Since 2026-09-18 Google serves the Gemini 2.5 models only to users who have used them before, and points new projects to 3.5 Flash-Lite o... Source ↗
limits reduced Groq: Groq shut down Llama 3.1 8B and Llama 3.3 70B on its free and developer plans. On the free plan Llama 3.1 8B had allowed 14,400 requests ... Source ↗
limits increased Composio: Composio's new pricing, for signups on or after 2026-08-15, raised the free plan from 20,000 tool calls a month (triggers counted inside ... Source ↗
pricing restructured Mistral AI: Mistral's free API moved to a $10 monthly allowance shared across Studio, the API and Vibe Code; when it is used up, usage can stop unles... Source ↗
restriction Mistral AI: Since 2026-08-05 Mistral's consumer terms exclude Mistral AI Studio and its API, which Mistral limits to business customers; API use fall... Source ↗
limits reduced Roboflow: Roboflow's free Public plan went from "$60/mo free credits" to "15 credits / month" between the Internet Archive's 2026-06-23 and 2026-07... Source ↗
limits reduced Groq: Groq shut down Llama 4 Scout and Qwen3 32B on its free and developer plans. Their model IDs were meta-llama/llama-4-scout-17b-16e-instruc... Source ↗
free tier removed Cerebras: Cerebras's free access became a time- and credit-bounded trial for new accounts: $5 in credits after adding a verified payment method, ex... Source ↗
restriction GitHub Copilot: GitHub limited Copilot Free and Copilot Student to auto model selection, so Free users can no longer choose which model answers. Source ↗

View all 513 pricing changes →

Open-Source Self-Hosted Alternatives

For each category, there's an open-source option you can run on your own hardware — no API limits, no vendor lock-in, complete data privacy.

Category Tool Details
LLM Inference Ollama Run Llama, Mistral, Gemma locally — no API limits, complete privacy. GPU recommended.
Experiment Tracking MLflow Apache-licensed ML lifecycle platform. Self-host for unlimited experiments and model registry.
Vector Database Chroma Open-source embedding database. Run locally in Python with zero configuration.
AI Observability Langfuse (self-hosted) Same Langfuse, no observation limits. Deploy via Docker or Kubernetes.
Data Labeling Label Studio Open-source annotation tool. Unlimited projects, all data types (text, image, audio, video).
AI Gateway LiteLLM Open-source proxy to 100+ LLMs. Unified API, load balancing, spend tracking. No request limits.
Compute Ollama + llama.cpp Run quantized models on consumer hardware. M-series Macs run 7B-13B models at usable speeds.

When You'll Outgrow Free Tiers

Most AI projects can prototype entirely on free tiers. Here's what typically hits limits first:

  1. GPU compute time (Kaggle 30 hrs/week) — fine-tuning and training sessions eat GPU hours fast
  2. LLM API rate limits (Groq 30 RPM) — production apps with concurrent users need higher throughput
  3. Vector storage (Pinecone 2 GB) — RAG applications with large document corpora
  4. Speech credits (Deepgram $200 one-time) — real-time transcription burns through credits quickly
  5. Observability volume (Langfuse 50K units/month) — high-traffic AI apps with detailed tracing

The good news: the open-source alternatives above have no limits. Ollama + Chroma + MLflow + Langfuse (self-hosted) gives you a complete local AI stack at zero ongoing cost.

Need general SaaS infrastructure too? See our Free Startup Stack for hosting, databases, auth, and more. Or search our full index of 1,498+ developer deals.

More Alternatives Guides

View all alternatives guides →