Genuinely free AI tooling — model APIs, local models, coding agents, parsing, retrieval and observability — with the catches stated rather than buried.
Before you rely on any of these
Free tiers change without notice — these were checked on 1 August 2026, so confirm the current limits before you build on one. And read the data terms: several free tiers train on what you send them, which makes them the wrong place for anything belonging to a client.
Model access
Where to get inference without a bill.
Inference on custom LPU chips — fast enough for voice and real-time work, where most providers are not.
Free: Long-standing free tier, stable for years. The default starting point for anything interactive.
Very high throughput inference, built for batch work.
Free: Among the most generous permanent quotas available — with a real catch: the top free tier requires opting into training on your data, so it must never see client material.
One API key routed across many providers, with automatic failover when one throttles.
Free: 20+ free models behind a single OpenAI-compatible endpoint. The least painful way to try models before committing.
Gemini with up to a million tokens of context.
Free: Free tier with the largest context here — the practical choice for reading a whole codebase or contract in one pass.
Models running at the edge, next to your other Cloudflare services.
Free: Generous request budgets, and barely mentioned compared to the others.
Running models yourself
Private by construction, and free permanently.
One command to run open models locally.
Free: Open source. Nothing leaves the machine, which settles most confidentiality questions before they are asked.
A desktop app over the same idea, with model discovery built in.
Free: Free to use, no account.
An offline assistant app — the local answer to a chat interface.
Free: Open source, fully offline.
Coding
Assistants that do not lock you to one vendor.
Terminal pair programmer where every edit lands as a git commit.
Free: Open source; you supply a model key. Strong on refactors and inherited code precisely because the history is legible.
Open-source VS Code and JetBrains extension that plugs into any model, including a local one.
Free: Open source. The route to an assistant with no vendor lock-in.
An agentic VS Code extension that reads, edits and runs.
Free: Open source; bring your own key.
Self-hosted code completion.
Free: Open source — the option when source code may not leave the network.
Documents into text a model can read
The least glamorous category, and where most retrieval projects quietly fail.
Prefix any URL and get clean markdown back.
Free: Free, no key needed for basic use. Removes a whole class of scraping code.
Document parsing that keeps tables and layout intact.
Free: Open source, from IBM Research. Far better on real-world PDFs than the usual libraries.
Speech and meetings
Transcription good enough to build on.
Whisper with word-level timestamps and speaker separation.
Free: Open source, runs locally. The timestamps are what make a transcript searchable rather than just readable.
Hosted speech-to-text with strong accuracy and diarisation.
Free: Free credits on sign-up, enough to evaluate properly.
Images
Including the two things most generators still cannot do.
Generates actual vector output — SVG logos and icons that stay editable.
Free: Free daily allowance. Very few tools produce vectors rather than pixels, which is why it belongs on any design shortlist.
Voice and video
Open models that hold their own against paid services.
A small open speech model with quality far above its size.
Free: Open weights. Runs locally, which also removes the per-character billing.
Video generation with a storyboard-shaped workflow.
Free: Free tier with monthly credits.
Building AI applications
Platforms you can self-host, so the free tier never ends.
An open-source platform for building LLM applications with retrieval built in.
Free: Self-host and it is free entirely. Deserves to be far better known than it is.
Workflow automation that drops into code when the visual nodes run out.
Free: Free when self-hosted; the price is that you now run a service.
Retrieval infrastructure
Where the embeddings live.
A fast vector database written in Rust.
Free: Free cloud tier, and open source if you would rather run it.
An embedded vector store — no server to run at all.
Free: Open source. The SQLite-shaped answer to vector search.
Vector search inside Postgres.
Free: Open source, and already available on Supabase — if you have Postgres, you have a vector database.
Seeing what your AI actually does
The step most teams skip, and the reason AI features fail quietly.
Tracing, cost tracking and evaluation for LLM applications.
Free: Open source and self-hostable, with a free cloud tier.
Tests prompts the way you would test code, so a change cannot silently regress.
Free: Open source.
Design, diagrams and research
The odds and ends that save real hours.
Turns written text into explanatory visuals.
Free: Free tier. Unusual, and genuinely useful for explaining a system.
Generates sitemaps and wireframes from a description of the business.
Free: Limited free tier — aimed squarely at agency proposal work.
Working out which of these your problem actually needs?
Book a call
