Free LLM APIs
Models offered at $0.00 per token. Free models carry no fee. 25 of 338 catalog models match as of July 2026.
- 25
- Models
- Free
- Input from /1M
- 1.0M
- Max context
- Jun 17, 2026
- Newest release
OpenKey currently routes 25 free models across 329 total — no cost, no card, same one-key setup as the paid catalog. The newest addition is North Mini Code (free) from Cohere, live June 17, 2026, with a 256K context window. Coverage spans small edge models like Liquid's LFM2.5-1.2B up through Nemotron 3 Ultra at a 1M-token context window. Only 8 of the 25 have published benchmark scores, so pick by verified data where it exists, and treat the rest as untested until you run your own eval.
Top picks
Qwen3 Coder 480B A35B (free)
The only free model with a coding-relevant benchmark on record: rank 54 (elo 1192) on design_arena codecategories. 1,048,576-token context handles full-repo prompts without chunking, and it's a 480B-A35B mixture-of-experts, not a toy checkpoint.
Nemotron 3 Ultra (free)
1,000,000-token context, released 2026-06-04, and the strongest general benchmark showing in the free tier: design_arena 3d rank 40 (elo 1213), well ahead of gpt-oss-120b's rank 84 (elo 1059) on its own benchmark.
gpt-oss-120b (free)
OpenAI's open-weight release (2025-08-05), 131,072-token context, with a measured design_arena gamedev rank 84 (elo 1059) — one of the few free models where you're not guessing at capability from the name alone.
All 25 models
Full catalog →North Mini Code (free)Cohere256KFreeFreectx 256K · inFreeoutFree
Nemotron 3 Ultra (free)NVIDIA1MFreeFreectx 1M · inFreeoutFree
Nemotron 3.5 Content Safety (free)NVIDIA128KFreeFreectx 128K · inFreeoutFree
Nemotron 3 Nano Omni (free)NVIDIA256KFreeFreectx 256K · inFreeoutFree
Laguna M.1 (free)poolside262KFreeFreectx 262K · inFreeoutFree
Laguna XS.2 (free)poolside262KFreeFreectx 262K · inFreeoutFree
Gemma 4 26B A4B (free)Google262KFreeFreectx 262K · inFreeoutFree
Gemma 4 31B (free)Google262KFreeFreectx 262K · inFreeoutFree
Lyria 3 Clip PreviewGoogle1.0MFreeFreectx 1.0M · inFreeoutFree
Lyria 3 Pro PreviewGoogle1.0MFreeFreectx 1.0M · inFreeoutFree
Nemotron 3 Super (free)NVIDIA1MFreeFreectx 1M · inFreeoutFree
Free Models RouterOpenRouter200KFreeFreectx 200K · inFreeoutFree
LFM2.5-1.2B-Instruct (free)Liquid AI33KFreeFreectx 33K · inFreeoutFree
LFM2.5-1.2B-Thinking (free)Liquid AI33KFreeFreectx 33K · inFreeoutFree
Nemotron 3 Nano 30B A3B (free)NVIDIA256KFreeFreectx 256K · inFreeoutFree
Nemotron Nano 12B 2 VL (free)NVIDIA128KFreeFreectx 128K · inFreeoutFree
Qwen3 Next 80B A3B Instruct (free)Qwen262KFreeFreectx 262K · inFreeoutFree
Nemotron Nano 9B V2 (free)NVIDIA128KFreeFreectx 128K · inFreeoutFree
gpt-oss-120b (free)OpenAI131KFreeFreectx 131K · inFreeoutFree
gpt-oss-20b (free)OpenAI131KFreeFreectx 131K · inFreeoutFree
Qwen3 Coder 480B A35B (free)Qwen1.0MFreeFreectx 1.0M · inFreeoutFree
Uncensored (free)Cognitive Computations33KFreeFreectx 33K · inFreeoutFree
Llama 3.3 70B Instruct (free)Meta AI131KFreeFreectx 131K · inFreeoutFree
Llama 3.2 3B Instruct (free)Meta AI131KFreeFreectx 131K · inFreeoutFree
Hermes 3 405B Instruct (free)Nous Research131KFreeFreectx 131K · inFreeoutFree
Prices per 1M tokens, flat 3% fee included. Hover or tap a price for the math.
How to choose
First tradeoff: benchmarked vs. unverified. Only 8 of 25 free models have any published score — Nemotron 3 Ultra, both Gemma 4 variants, Nemotron 3 Super, Nemotron 3 Nano 30B, both gpt-oss sizes, and Qwen3 Coder. Everything else (including Hermes 3 405B and Llama 3.3 70B) is a known architecture running blind on OpenKey's benchmark data. Second: context window varies by 32x in this list — Liquid's LFM2.5 models cap at 32,768 tokens while Nemotron 3 Ultra and Nemotron 3 Super go to 1,000,000. Match window to task, not to hype. Third: modality. Most free models are text-to-text, but Gemma 4 26B/31B and Nemotron 3 Nano Omni add image/video input, and Lyria 3 (Clip/Pro Preview) does text+image to text+audio — pick by what you're actually feeding the model.
Questions
- Are these free models actually free on OpenKey, or is there a catch?
- They're free at the provider level — input and output are both $0.00 per million tokens for all 25 models. OpenKey's 3% fee applies to provider list price, and 3% of zero is zero. You still use the same single API key as the paid catalog; no separate signup.
- Which free model has the best benchmark score right now?
- Nemotron 3 Ultra (free) posts the highest elo among free models with data: 1213 on design_arena 3d, ranked 40th. That's ahead of Qwen3 Coder (elo 1192, rank 54) and both gpt-oss models (elo 1059 and 978). Only 8 of 25 free models have any benchmark on record, so this comparison is limited to that subset.
- What's the newest free model on OpenKey?
- North Mini Code (free) from Cohere, created 2026-06-17, with a 256,000-token context window. It has no published benchmark yet in OpenKey's data, so treat it as unverified until scores land — same caveat as most of this free tier.