1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121 | # Free AI Resources
**Requirements:** none — this is a plain markdown guide, nothing to install, full walkthrough in [SETUP.md](SETUP.md).
A curated list of genuinely free ways to use AI in 2026 — frontier chat models, local models you run yourself, image and audio generation, developer tooling, and learning resources. No affiliate links, no paid placements. Where "free" comes with a catch (rate limits, local compute requirements, feature gating), that catch is stated plainly.
## Why this exists
Most "free AI tools" lists are stale within a month or are secretly monetized content. This one is meant to be maintained by hand and to say "check current pricing" whenever a vendor's free tier is known to shift often, instead of pretending a number is stable when it isn't.
## How to use this
Read by category. Each entry has: what it is, what "free" actually means for it, and where to get it. Nothing here requires a credit card unless explicitly noted.
---
## Frontier chat models (hosted, free tier)
These are the big commercial labs' free-to-use interfaces. All of them rate-limit free usage and reserve their strongest models or highest usage caps for paid tiers.
- **Claude (Anthropic)** — claude.ai offers free-tier access to a Claude model with daily/session message limits that reset periodically. Model availability and limits on the free tier change over time — check claude.ai for current terms. Claude Code (the CLI) also has usage included on paid Claude subscriptions; check current terms for what's free vs. subscription-gated.
- **ChatGPT (OpenAI)** — chatgpt.com offers a free tier with a capped, rotating model and usage limits. Free-tier model access and limits change frequently — check current terms at the time you sign up.
- **Gemini (Google)** — gemini.google.com offers a free tier, generally with the most generous integration into Google's own products (Docs, Gmail, etc.) if you already have a Google account. Limits and model tier vary — check current terms.
**Honest note:** none of these free tiers are stable long-term promises. Vendors adjust limits and default models often. Treat the above as "there is a free tier," not "here is exactly what you get."
---
## Free local models (run on your own hardware via Ollama)
Running a model locally means no per-message cost and no vendor rate limit — but you need adequate local compute (RAM/VRAM), and quality is generally below the frontier hosted models unless you're running a very large local model on serious hardware.
Install [Ollama](https://ollama.com) (free, open source, macOS/Linux/Windows), then pull any of these:
- **Llama family (Meta)** — `ollama pull llama3.2` (and newer Llama releases as Ollama adds them). Open-weight, runs from a few GB (small distills) up to much larger sizes if your hardware supports it.
- **Qwen family (Alibaba)** — `ollama pull qwen2.5` (and newer). Strong performance-per-parameter, widely used for local coding and general chat.
- **Gemma family (Google)** — `ollama pull gemma2` (and newer). Google's open-weight line, tuned for efficient local inference.
- **DeepSeek-R1 distills** — `ollama pull deepseek-r1` (distilled sizes, e.g. 7B/8B/14B, run on consumer hardware; the full-size R1 model needs a lot more RAM/VRAM). Reasoning-focused, open-weight.
**What "free" means here:** the model weights and Ollama itself cost nothing. The real cost is your own compute (a laptop or desktop with enough RAM, ideally a GPU or Apple Silicon unified memory) and the time to run inference locally. There is no per-message bill and no vendor account required.
---
## Free image generation
- **Local, via Hugging Face Diffusers / ComfyUI / Automatic1111** — open-weight image models (e.g. Stable Diffusion variants, FLUX.1-schnell) can be run entirely locally, free, with no usage cap other than your own hardware and time. Requires a GPU with enough VRAM for reasonable speed; CPU-only is possible but slow.
- **Hugging Face Spaces** — many community-hosted Spaces let you run image generation in a browser for free on shared/community GPU compute, with queue times and usage limits that vary by Space. No install required, but availability is not guaranteed and can be slow at peak times.
- **Hosted free tiers (Bing Image Creator / Google, etc.)** — some hosted consumer products bundle limited free image generation into a free account. These change often — check current terms before relying on one.
---
## Hugging Face (models, datasets, Spaces)
[huggingface.co](https://huggingface.co) is free to use for browsing, downloading, and hosting:
- **Models** — tens of thousands of open-weight models (text, image, audio, multimodal) are free to download and run locally, subject to each model's own license (some are fully open, some restrict commercial use — read the model card).
- **Datasets** — a large public dataset hub, free to download, useful for fine-tuning or evaluation work.
- **Spaces** — free hosting for small demo apps (Gradio/Streamlit), with a free CPU tier and paid GPU tiers if you need more power. Community Spaces built by others are also free to use.
- **Transformers / Diffusers libraries** — open source Python libraries for running models locally, free, well documented.
---
## Free embeddings
- **Local via sentence-transformers (Hugging Face)** — open-weight embedding models (e.g. the `sentence-transformers` model family) run fully free and locally via the `sentence-transformers` Python package, no API key or per-call cost.
- **Ollama embedding models** — some Ollama-distributed models support embedding generation locally (check a given model's Ollama page for embedding support).
- **Hosted free tiers** — some vendors offer a limited free monthly quota for embedding API calls; these are usage-capped and vendor terms change — check current pricing before depending on one for production volume.
---
## Free speech-to-text (STT)
- **whisper.cpp (local)** — a free, open-source, dependency-light C/C++ port of OpenAI's Whisper model that runs entirely on your own CPU/GPU, no API key, no per-minute cost. Runs well on Apple Silicon and most modern laptops. Get it from [github.com/ggml-org/whisper.cpp](https://github.com/ggml-org/whisper.cpp).
- **faster-whisper / openai-whisper (local, Python)** — the original Whisper model and faster reimplementations, also free to run locally with a Python environment.
**What "free" means here:** no vendor account, no per-minute charge, no internet connection required once the model is downloaded. Cost is local compute time.
---
## Free text-to-speech (TTS)
- **Chatterbox (MIT licensed)** — an open-source TTS model released under the MIT license, free to run locally and free to use commercially (subject to the MIT license terms — read it, but MIT is one of the most permissive licenses available). Runs on local hardware.
- **mlx-audio (Apple Silicon)** — an MLX-based audio/TTS toolkit optimized to run efficiently on Apple Silicon Macs, free and open source, no cloud dependency once set up locally.
**What "free" means here:** these are self-hosted — you provide the compute, the software costs nothing, and there's no per-character or per-minute vendor charge.
---
## Free developer tools
- **Claude Code (Anthropic)** — usage is tied to your Claude account/subscription; check current terms for what free-tier or included usage looks like, since this changes over time.
- **Cursor** — an AI-assisted code editor with a free tier that includes a limited number of AI-assisted completions/requests per period. Free-tier limits change often — check current pricing on cursor.com before relying on a specific quota.
- **Windsurf** — similarly offers a free tier with usage limits that have changed multiple times since launch — check current pricing on their site rather than trusting any number written down elsewhere, including this one.
- **Ollama + any local model above** — a completely free, no-account-required way to get AI code assistance locally if you pair a local coding-tuned model (e.g. a Qwen or DeepSeek coder variant) with an editor extension that talks to a local Ollama server.
**General rule for this section:** IDE/editor AI free tiers are the most volatile category in this entire list. Vendors change limits, models, and pricing structures frequently. Always verify current terms directly on the vendor's site before making a decision based on a specific number.
---
## Free learning resources
- **fast.ai** — Jeremy Howard and Rachel Thomas's free, practical deep learning courses (Practical Deep Learning for Coders and others), no cost, no account required to watch/read. [fast.ai](https://www.fast.ai)
- **Andrej Karpathy's "Neural Networks: Zero to Hero"** — a free video series building neural networks and language models from scratch, code-along style, available free on YouTube.
- **Hugging Face course** — a free course covering Transformers, NLP, and the Hugging Face ecosystem, available at [huggingface.co/course](https://huggingface.co/learn).
These three are stable, well-established free resources maintained by their authors/organizations over multiple years — the most durable entries on this list.
---
## Make it your own
This is a plain markdown file — there's no build step. To adapt it:
1. Fork or clone this repo.
2. Edit `README.md` directly. Add categories, remove ones you don't use, or narrow it to your own stack.
3. If you maintain your own copy, re-check every "check current pricing" entry periodically — that phrase exists because those specific items are known to move, not as a hedge to avoid doing the work.
4. If you find an entry that's gone stale or wrong, fix it and note the date you verified it, so future readers know how fresh the information is.
## A note on accuracy
Every claim in this file is written to be checkable against the vendor's own current documentation or the tool's own repository. Nothing here is an invented benchmark, discount, or performance number. Where free-tier specifics (limits, exact models included) are known to change frequently, this file says so instead of stating a number that will go stale.
|