Updated November 2026. Free limits re-checked against each provider's docs on November 11, 2026.
Free AI Models for Coding in Cursor, VS Code + Claude Code (2026)
You do not need a paid API plan to get serious AI coding help in 2026. The best free options right now are GPT OSS 120B on Groq (1,000 free requests per day through an OpenAI-compatible endpoint), Cohere's North Mini Code on OpenRouter (a free agentic coding model with a 256K context window), Qwen2.5-Coder running locally in Ollama (the local autocomplete model Continue recommends), and self-hosted open weights like GLM 4.6 or DeepSeek V3.2 through Ollama or vLLM, which cost nothing and keep your code on your machine.
What changed since July: three of our earlier picks stopped being free. Mistral's free mode no longer includes API access, so Codestral needs billing turned on. Groq's free tier no longer lists Llama 3.3 70B. OpenRouter no longer offers a free DeepSeek model. If a free coding setup broke for you this autumn, that is probably why.
All four plug into the editors developers actually use. Cursor accepts any OpenAI-compatible base URL in its model settings. VS Code gets there through extensions like Continue. Claude Code runs on Anthropic models, but free models slot in alongside it through MCP servers and router setups.
This guide covers each option, the exact free limits, how to wire them into your editor, and where free tiers break down on real work. We maintain a live directory of free models at /free-models, filterable by Code and re-checked against each provider's docs, with the rate limits and the catch for every entry, so the specifics below stay verifiable even as providers change their terms.
What are the best free AI models for coding in 2026?
Four options cover almost every situation: a fast hosted generalist, a hosted code specialist, a local autocomplete model, and a self-hosted setup for private code.
| Model | Where it runs | Free limits | Best for | The catch |
|---|---|---|---|---|
| GPT OSS 120B | Groq's free tier | 30 requests per minute, 1,000 per day, 8K tokens per minute | Fast chat, refactors, code review | The 8K tokens-per-minute cap limits how much code you can paste per request |
North Mini Code (:free) | OpenRouter | 20 requests per minute, 50 per day (1,000 per day after buying 10+ credits) | Agentic coding, long files (256K context) | 50 requests a day runs out fast in an agent loop |
| Qwen2.5-Coder 1.5B / 7B | Local via Ollama | None | Tab autocomplete in VS Code | Small models, so keep them on autocomplete, not hard reasoning |
| DeepSeek V3.2 / GLM 4.6 | Self-hosted via Ollama or vLLM | None, free forever | Proprietary codebases, offline work | You need serious local hardware for the larger weights |
For a broader directory of AI models, paid and free, hosted and open-weights, browse our full model directory.
GPT OSS 120B on Groq is the speed play. Groq's custom inference hardware returns tokens fast enough that chat feels instant, and the free tier allows 30 requests per minute and 1,000 per day for openai/gpt-oss-120b, per Groq's rate-limits page. The catch is 8,000 tokens per minute, so paste the function you are working on, not the whole repo. The endpoint is OpenAI-compatible at https://api.groq.com/openai/v1, which means nearly every editor and tool that speaks the OpenAI API can use it with a one-line base URL change.
North Mini Code on OpenRouter is the code specialist. Cohere describes it as its first agentic coding model: a sparse mixture-of-experts model with 30B total and 3B active parameters and a 256K context window. One OpenRouter key and the OpenAI-compatible endpoint at https://openrouter.ai/api/v1 give you cohere/north-mini-code:free, plus the other models tagged :free. Free use is capped at 20 requests per minute and 50 per day, rising to 1,000 per day once you have bought at least 10 credits.
Qwen2.5-Coder in Ollama replaces Codestral for autocomplete. Codestral is still the hosted model Continue recommends for tab completion, but Mistral's free mode no longer includes API access, so it now costs money. Continue's local recommendation is qwen2.5-coder:1.5b (or the 7B version on a stronger machine). It supports fill-in-the-middle (FIM) completion, which is what makes editor autocomplete feel native: the model sees the code before and after your cursor. Because it runs on your machine, there is no rate limit at all.
Self-hosting DeepSeek V3.2 or GLM 4.6 is the privacy play. Both are open-weight releases. Pull them with ollama run for a laptop-friendly setup or serve them with vLLM on a GPU box for team throughput. There is no rate limit, no usage meter, and no terms-of-service change that can take the model away. Your code never leaves your network.
Key takeaway: hosted free tiers (Groq, OpenRouter) win on model quality per minute of setup, while local and self-hosted open weights win on privacy and predictability. Most developers end up running one of each.
How do free coding model tiers actually differ?
The gap between free models is much wider than most comparison posts admit. On agentic coding benchmarks like SWE-bench, scores across free-tier models span a huge range: large open-weight models resolve a meaningful share of real GitHub issues, while small local models that fit comfortably on a laptop resolve far fewer. The tiers are real.
That leaves you with one honest tradeoff: quality versus privacy and cost. The strongest free models are hosted, which means your prompts and code snippets transit someone else's servers under someone else's data policy. The most private models run locally, but the versions that fit consumer hardware are quantized or smaller, and it shows on multi-file refactors.
The practical read:
- Autocomplete tolerates smaller models well. A local Qwen2.5-Coder covers it with no limits.
- Chat about code wants a large model. GPT OSS 120B on Groq is the free sweet spot.
- Agentic work (multi-step edits, running tests, fixing failures) is where free tiers strain. North Mini Code on OpenRouter is built for it, but 50 free requests a day interrupt long loops.
How do I use free AI models in Cursor?
Cursor supports custom OpenAI-compatible endpoints, so any of the hosted free tiers above can power its chat. The setup takes about two minutes:
- Get a key. For Groq, create one at console.groq.com. For OpenRouter, at openrouter.ai. Both are free.
- Open Cursor Settings, then Models.
- In the OpenAI API section, enable Override OpenAI Base URL and paste the endpoint:
https://api.groq.com/openai/v1for Groq orhttps://openrouter.ai/api/v1for OpenRouter. - Paste your API key in the key field and verify.
- Add the model name Cursor should request:
openai/gpt-oss-120bfor Groq orcohere/north-mini-code:freefor OpenRouter. - Select that model in the chat model picker and test with a small prompt.
Two things to know. First, overriding the base URL routes Cursor's chat requests through your custom endpoint, but Cursor's own Tab autocomplete uses Cursor's proprietary models and is not swappable this way. Second, if a model name errors out, check the provider's model list, since free model IDs change more often than paid ones. Our /free-models directory tracks the current IDs and limits.
How do I use free AI models in VS Code?
Plain VS Code does not ship a bring-your-own-model chat, so the standard route is an extension. Continue is the most established open-source option and speaks both OpenAI-compatible APIs and local Ollama:
- Install the Continue extension from the VS Code marketplace.
- Open Continue's configuration (the gear icon in the Continue panel).
- Add a model block pointing at your provider. For Groq, set the provider to an OpenAI-compatible endpoint with base URL
https://api.groq.com/openai/v1, modelopenai/gpt-oss-120b, and your key. For a local model, set the provider to Ollama and the model to the tag you pulled. - For autocomplete, configure a separate tab autocomplete model. Run
ollama run qwen2.5-coder:1.5b, then add a model block with providerollama, modelqwen2.5-coder:1.5band theautocompleterole. If you are willing to pay, Continue's docs still list Codestral as their recommended autocomplete model. - Reload the window and confirm both chat and autocomplete respond.
This split setup, local Qwen2.5-Coder for autocomplete plus GPT OSS 120B on Groq for chat, is the strongest fully free VS Code configuration we know of. Both halves cost nothing and neither requires a credit card.
Can Claude Code use free models?
Not directly, and it is worth being precise about this. Claude Code is Anthropic's agentic coding tool and it requires Anthropic models to run. You cannot point it at Groq or OpenRouter and expect the agent loop to work.
What teams actually do is pair it: Claude Code drives the agentic work, and free models handle the high-volume, low-stakes calls around it. Common patterns include MCP servers that expose a free model as a tool Claude Code can call (for bulk summarization, test data generation, or second-opinion reviews), and router setups where scripted pipeline steps hit Groq or a local Ollama model while the reasoning-heavy steps stay on Claude. The community tooling for this moves fast; our roundup of the best Claude Code GitHub repos covers the MCP servers worth starting from.
If your goal is a completely free agentic setup, the honest answer is that you want an open-source harness (Continue, Aider, or similar) driving a free model like GPT OSS 120B or a self-hosted open-weight model rather than Claude Code with a workaround bolted on.
Why do teams self-host free models for proprietary code?
Because with a hosted free tier, your code is the payment. Every completion request sends surrounding file context to the provider, and free tiers frequently have broader data retention or training rights than paid tiers. For an open-source side project that is a fine trade. For a proprietary codebase with customer data models, API keys in fixtures, or unreleased product logic, it is often a compliance non-starter.
Self-hosting closes that door completely:
- Ollama is the low-friction path.
ollama run qwen2.5-coder:7b(or another model sized to your RAM) gives you a local OpenAI-compatible server that Continue and other editor extensions connect to out of the box. - vLLM is the team path. Serve GLM 4.6 or DeepSeek V3.2 on a shared GPU server and every developer points their editor at one internal endpoint. Throughput is high enough for a whole team, and the marginal cost per request is zero.
The tradeoff from earlier applies: local models sized for a laptop are weaker than hosted frontier models. But for autocomplete and routine chat over sensitive code, a local open-weight model is the only option that requires zero trust in a third party.
Key takeaway: if the code is proprietary, self-host. An open-weight model behind your own firewall beats a stronger hosted model you are not allowed to send code to.
When does free break down?
Free tiers fail in predictable ways, and it pays to know them before you are mid-task:
- Rate limits mid-refactor. Agentic edits burn requests fast. A multi-file refactor with test runs can fire dozens of calls in a minute, which hits Groq's 30 requests per minute fast, eats a large chunk of its 1,000 daily requests in one session, and can use up OpenRouter's 50 free daily requests before lunch. The failure mode is a stalled agent halfway through a change, leaving your branch in a half-edited state.
- Context limits. Free variants often cap context below the paid version of the same model. Large monorepo questions that need 100K+ tokens of context get silently truncated, and the model answers confidently about the fraction of the codebase it actually saw.
- Queue deprioritization. OpenRouter's
:freetraffic sits behind paid traffic. At peak hours, latency spikes or requests fail, which is survivable for chat and painful for autocomplete. - Silent policy changes. Free limits are not contracts. Providers adjust them without much notice, which is exactly why we track current limits per model at /free-models instead of hardcoding them into docs.
None of this means free models are a toy. It means free models are excellent for evaluation, side projects, autocomplete, and cost-capping specific workloads, and unreliable as the sole engine behind production delivery deadlines.
Ship faster with AI-native engineers
Here is the pattern we see across teams: picking the model is the easy 10 percent. The hard 90 percent is engineers who know how to drive these tools, when to trust an agentic edit and when to read every line, how to structure a codebase so AI tooling works well in it, and how to build the review discipline that keeps velocity from turning into incident volume.
That is a hiring problem, not a tooling problem. At AY Automate we place AI-native engineers, developers who work fluently with Cursor, Claude Code, and self-hosted model stacks, directly into product teams. If your roadmap is moving slower than your tooling budget suggests it should, our engineer placement service matches you with engineers who ship with these workflows on day one.
Start with the free stack in this guide, prove the workflow on your own codebase, and when you are ready to scale it across a team, book a consultation or read more about how engineer placement works. The models are free. The velocity comes from the people using them.
FAQ
Can I use Cursor with a free AI model?
Yes, Cursor's chat can run on free models through its custom OpenAI base URL setting. Point the override at https://api.groq.com/openai/v1 or https://openrouter.ai/api/v1, add your free key, and add the model name. Cursor's Tab autocomplete, however, uses Cursor's own models and cannot be swapped for a free external one.
What is the best free coding LLM?
As of November 2026, GPT OSS 120B on Groq is the best free hosted model for coding chat (1,000 requests a day), and Cohere's North Mini Code on OpenRouter is the best free model built for agentic coding. For autocomplete, run Qwen2.5-Coder locally in Ollama. For private code, self-host open weights such as DeepSeek V3.2 or GLM 4.6.
Is DeepSeek free for coding?
Not on OpenRouter anymore: it no longer offers a free DeepSeek model (checked November 11, 2026). Smaller hosts still have tight free tiers, such as DeepSeek V3.1 on SambaNova at 20 requests a day, which is too few for real coding work. DeepSeek's open-weight releases like V3.2 can still be self-hosted with Ollama or vLLM at no license cost, but the large versions need serious GPU hardware.
Do free AI coding models train on my code?
Sometimes, and you should assume they might unless the provider's data policy says otherwise. Free tiers commonly carry broader data usage rights than paid tiers. If that risk is unacceptable, self-host an open-weight model so no code leaves your machine.
Is Codestral still free?
Not through Mistral's own API. Mistral's usage-limits page says free mode covers a monthly budget for trying Vibe and Studio but does not include API access: creating API keys and calling the API require billing and pay-as-you-go. For free autocomplete, use a local Qwen2.5-Coder model in Ollama instead.
Can I run a free coding model completely offline?
Yes. Open-weight models like DeepSeek V3.2 and GLM 4.6 run locally through Ollama or vLLM with no network access required after the initial download. Quality depends on how large a variant your hardware can hold, but the workflow itself is fully offline and private.
How many requests per day does Groq's free tier allow?
Groq's free tier allows 1,000 requests per day for GPT OSS 120B, along with 30 requests per minute and 8,000 tokens per minute, per the limits published at console.groq.com/docs/rate-limits. Llama 3.3 70B is no longer on the free list. The endpoint is OpenAI-compatible, so the same key works in Cursor, Continue, and most other editor integrations without custom code.
For related guides in this cluster: Best free AI models for n8n covers no-cost agent workflows, free AI models for commercial use covers license-safe shipping, and best local LLM tools 2026 covers self-hosted inference when rate limits are a constraint.
Once you are running free models in a coding workflow, the best AI agent observability tools guide covers cost-per-request tracking and latency monitoring.
Sources (checked November 11-13, 2026): console.groq.com/docs/rate-limits, openrouter.ai/api/v1/models, docs.mistral.ai/admin/billing-usage/usage-limits, docs.continue.dev (autocomplete), ollama.com, docs.vllm.ai
Book a Free Strategy Call
Building this in production?
Walid runs a 30-min call to map your AI engineering team. Free, no slides.

Ex-IBM AI engineer and enterprise architect. Adel owns the technical architecture behind every automation and AI agent system AY Automate ships.
