"Free" in AI infrastructure usually means a trial credit that evaporates in a week. But several providers now run genuinely free tiers with no expiry: rate-limited rather than credit-limited. If you pick correctly, you can run Vexi indefinitely without entering a card.
Vexi flags these providers with a FREE badge. Here's the honest breakdown.
Groq: the fast one
Groq runs open models on custom inference hardware, and the speed difference is not subtle. Responses arrive faster than you can read them, which changes how an agent feels: you stop waiting and start iterating.
The tradeoff is model ceiling. You're using Llama and similar open weights, not a frontier model. For refactors, boilerplate, test generation, and "explain this function," that's plenty. For subtle architectural reasoning across a large codebase, you'll feel the gap.
Free tier is rate-limited per minute and per day. Solo work rarely hits it; a long agentic session sometimes will.
Google Gemini: the generous one
Gemini's free tier through AI Studio is the most usable of the group for serious work. Flash models are quick, the context window is enormous, and the daily request allowance is high enough that most developers never reach it.
If you want one free provider and don't want to think about it further, start here.
Cerebras: the very fast one
Same premise as Groq: custom silicon, open models, extremely low latency. Worth having configured as an alternative when Groq is rate-limiting you, since the two rarely throttle at the same time.
DeepSeek: the cheap reasoner
DeepSeek isn't free forever, but it's close enough to matter: pricing is low enough that ordinary daily use costs cents, and new accounts get credit to start. Its reasoning quality on code punches well above its price.
If you've exhausted the truly free options and want the best value per dollar, this is usually the answer.
GLM, Qwen, Kimi, MiniMax: the regional four
Zhipu's GLM, Alibaba's Qwen, Moonshot's Kimi, and MiniMax all offer free tiers, and all four are strong on code. GLM-4-Flash in particular is fast and free.
Two practical caveats. First, latency from outside Asia can be noticeably higher. Second, and this trips people up, several of these run separate Chinese and international platforms with separate key systems. A key from the wrong console will fail authentication even though it looks valid. Vexi's provider links point at the console matching the endpoint it actually calls, which avoids this.
A practical setup
You don't have to choose one. Configure a fast free provider for everyday work and keep a stronger paid key for the hard problems:
# everyday: fast and free
vexi config set-key gsk_...
# switch when you hit something genuinely hard
vexi config set-key sk-ant-...Use /usage to watch what you're actually spending. Known free-tier models report as $0, so the number you see is real cost, not theoretical token accounting.
Where free stops working
Being straight about the limits: free tiers throttle under sustained agentic load. An agent that reads twenty files, plans an edit, applies it, and re-reads to verify burns requests fast. Long autonomous sessions on a free tier will hit a wall.
The failure is graceful (you get a rate-limit error, wait, and continue), but if you're doing multi-hour agentic work daily, a paid key for a few dollars a month removes the friction entirely. Start free, and upgrade only when the limits actually annoy you.