Free reasoning models today
The table lists free models with reasoning in the name, for example GPT-OSS, R1, QwQ, Thinking, Reasoning and Magistral. The data comes from the live catalog. Many other models, such as Gemini Flash, also think, but their names do not show it.
| Model | Free on | Context | Free limits (best offer) | Tools | Vision |
|---|---|---|---|---|---|
| Gemini 3.8 Live Extended Thinking | Google AI Studio | - | 10 req/min 20 req/day 250,000 tokens/min ~3M | Yes | Yes |
| GPT-OSS 20B | Api.Airforce, Cloudflare Workers AI, Groq +5 more | 131K | 30 req/min 1,000 req/day 8,000 tokens/min 200,000 tokens/day ~6M | Yes | No |
| GPT-OSS 120B | Cloudflare Workers AI, ElectronHub, Groq +2 more | 131K | 30 req/min 1,000 req/day 8,000 tokens/min 200,000 tokens/day ~6M | Yes | No |
| GPT-OSS Safeguard 20B | Groq, LLMTR | 131K | 30 req/min 1,000 req/day 8,000 tokens/min 200,000 tokens/day ~6M | Yes | No |
| OpenAI: GPT OSS 20B | ElectronHub | 128K | $0.25/week shared free credits | No | No |
| DeepSeek V4 Flash Thinking | ElectronHub | - | $0.25/week shared recurring credit | No | No |
| Qwen3 235B A22B Thinking 2507 | ModelScope | 262K | 200 req/day Daily login: 200 credits + 50 for linked Alibaba Cloud; 24h expiry; shared; noncommercial experience | No | No |
| GPT-OSS 120B (via Groq) | DreamPrompting | 131K | 100 req/min 5,000 req/day 500,000 tokens/day 500K tokens/day · 5,000 req/day | No | No |
| Qwen3 Next 80B A3B Thinking | ModelScope | 262K | 200 req/day Daily login: 200 credits + 50 for linked Alibaba Cloud; 24h expiry; shared; noncommercial experience | No | No |
| QwQ 32B | Cloudflare Workers AI | 131K | ~10-20M | Yes | No |
| Nemotron 3 Nano Omni Reasoning | BlockRun, Kilo Gateway, Requesty | 262K | 200 req/day 200 rpd | No | Yes |
| Nemotron 3 Nano Omni 30B A3B Reasoning | NVIDIA NIM, Waterfall | 262K | 40 req/min 40 RPM | No | Yes |
| Command A Reasoning | Cohere | 256K | 20 req/min 33 req/day ~1-2M | Yes | No |
| Nemotron 3 Nano 30B Reasoning | OpenRouter | 262K | 20 req/min 50 req/day 50 req/day account-wide (1,000 after a $10 purchase) | No | No |
| Nemotron 3 Nano Omni | ElectronHub | 256K | $0.25/week shared recurring credit | Yes | Yes |
| GPT-OSS 20B (via Groq) | DreamPrompting | 131K | 100 req/min 5,000 req/day 500,000 tokens/day 500K tokens/day · 5,000 req/day | No | No |
| DeepSeek R1 Distill Qwen 32B | Cloudflare Workers AI | 131K | ~3-5M | Yes | No |
| R1 Distill Qwen 32B | ElectronHub | 131K | $0.25/week shared recurring credit | Yes | No |
Live data from the FreeLLMAPI catalog (version 2026.11.19, checked 19 November 2026). Free limits change often. The catalog is checked every day.
GPT-OSS 120B is served free by several providers, so failover has many options. QwQ 32B and DeepSeek R1 Distill Qwen 32B run on Cloudflare Workers AI. Qwen3 Thinking models have a 262K-token context on ModelScope.
Thinking tokens use your output budget
This is the most common problem with reasoning models. The model writes its thinking first. These thinking tokens count against the output limit, the same as the answer.
- Google says: the output limit on Gemini includes thought tokens. If the model reaches the limit while it thinks, the output can be cut or empty. The thinking tokens are still counted.
- Groq says: the default of 1,024 completion tokens can be too low for complex reasoning.
- Free limits: thinking tokens also count against tokens-per-minute and tokens-per-day limits. A reasoning answer can use many more tokens than a direct answer to the same question.
max_tokens on a reasoning model. Start at 4,000 or more. If finish_reason is length and the content is empty, the model used the full budget to think.Step 1: Install FreeLLMAPI and add keys
FreeLLMAPI is open source. It runs on your computer in Docker. Your provider keys stay on your computer.
Install (macOS or Linux)
curl -fsSL https://freellmapi.co/install.sh | bashOpen the dashboard at http://localhost:3001. Add Groq, Cloudflare and Google AI Studio keys on the Keys page. Each one serves free reasoning models. Copy your unified key from the same page.
Step 2: Call a reasoning model
curl
curl http://localhost:3001/v1/chat/completions \
-H "Authorization: Bearer <unified-key>" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-oss-120b",
"max_tokens": 4000,
"messages": [{"role": "user", "content": "A bat and a ball cost 1.10 in total. The bat costs 1.00 more than the ball. What does the ball cost?"}]
}'Python: check for a cut answer and remove thinking text
import re
from openai import OpenAI
client = OpenAI(base_url="http://localhost:3001/v1", api_key="<unified-key>")
resp = client.chat.completions.create(
model="openai/gpt-oss-120b", # or "auto:smart" to rank by intelligence
max_tokens=4000,
messages=[{"role": "user", "content": "Plan a 3-step database migration with no downtime."}],
)
choice = resp.choices[0]
text = choice.message.content or ""
text = re.sub(r"<think>.*?</think>", "", text, flags=re.S).strip() # some models return visible thinking
if choice.finish_reason == "length" and not text:
print("The model used the full budget to think. Raise max_tokens.")
else:
print(text)
print("tokens used:", resp.usage.completion_tokens if resp.usage else "unknown")Some providers return the thinking inside <think> tags in the content. Others hide it, or return it in a separate field. Remove the tags before you show the answer to users.
Control the amount of thinking
Providers have their own settings for the thinking length. Groq accepts reasoning_effort with low, medium or high on GPT-OSS models. Gemini 3 models use a thinking_level setting. The FreeLLMAPI docs do not describe how these fields pass through the gateway, so test them on your install first.
The native Gemini endpoint of FreeLLMAPI, /v1beta/models/{model}:generateContent, translates thinking budgets. Use it if you need control over Gemini thinking.
| Task | Use a reasoning model? | Why |
|---|---|---|
| Math, logic puzzles, planning | Yes | Step-by-step thinking reduces errors |
| Hard bugs and algorithm design | Yes | The model checks its own steps |
| Chat, short answers, translation | No | Thinking adds time and tokens with little gain |
| Agents with many tool calls | Sometimes | Each round pays the thinking cost again |
Prompts and routing for reasoning tasks
- State the goal and the rules: give all facts, limits and the expected result in the prompt. A reasoning model cannot ask you for missing facts during its thinking.
- Set the answer format: ask for the final answer in a fixed form, for example one number or a JSON object. Then your code can read it without the thinking text.
- Give test cases: for code, add example inputs and outputs. The model can check its answer against them.
- Route by task: use
model: "auto:smart"to rank models by intelligence for hard tasks. Useauto:fastfor easy tasks, so that you do not pay the thinking cost. - One hard question for each request: several hard questions in one request share one output budget. Split them into separate requests.
The auto:* strategies rank every enabled model and ignore your chain order. To keep reasoning requests on reasoning models only, pin a model id. You can also make a named fallback chain with only reasoning models. Then call it with auto:<profile-name>.
Troubleshooting
| Problem | Cause | Fix |
|---|---|---|
Empty content, finish_reason: "length" | Thinking used the full output budget | Raise max_tokens to 4,000 or more |
<think> text in the answer | The provider returns visible thinking | Remove the tags in your code |
| Long explanation for a short question | Some reasoning models explain their thought process | Ask for "the final answer only", or use a non-reasoning model |
| Error 429 after a few requests | Thinking tokens used the tokens-per-minute limit | Add more provider keys, or use reasoning only for hard tasks |
| Slow answers | The model writes many thinking tokens | Use auto:fast for simple tasks |
Frequently asked questions
Is there a free reasoning model API?
Yes. Groq, Cloudflare, NVIDIA, OVH, OpenRouter and other providers give reasoning models free, such as GPT-OSS 120B, QwQ 32B and DeepSeek R1 Distill. FreeLLMAPI puts them behind one OpenAI-compatible key.
Do thinking tokens count as output tokens?
Yes. Google says that the Gemini output limit includes thought tokens, and that thinking tokens are counted. On free tiers they also use your tokens-per-minute limit.
Why does my reasoning model return an empty answer?
The model used the full max_tokens budget to think, so no tokens were left for the answer. Raise max_tokens to 4,000 or more.
Is DeepSeek R1 free through an API?
DeepSeek R1 Distill Qwen 32B is in the free catalog on Cloudflare Workers AI. The table on this page shows the providers that serve it today.
Which free reasoning model is best?
GPT-OSS 120B is a good default because several providers serve it free. Use the table on this page, which sorts models by intelligence rank from the live catalog.
Related guides
Sources
- FreeLLMAPI docs: Routing strategies and native Gemini API (checked 20 November 2026)
- Google AI Studio docs: Thinking (checked 20 November 2026)
- Gemini Developer API pricing (checked 20 November 2026)
- Groq docs: Reasoning (checked 20 November 2026)