Free models, with automatic quality gates.
Flow AI continuously discovers free-tier hosted models, runs each through a live canary — real tool-calling, structured output, response quality — and promotes only the ones that pass into the default routing pool. When a free model can genuinely do your task, auto routes to it and the tokens cost $0. When it can't, routing escalates to a paid model that finishes the job. No flag to set, nothing to configure.
Currently promoted (25)
Live list — models enter and leave as they pass or fail the canary. Refreshes every few minutes.
| Model | Context | Verified capabilities | Token cost |
|---|---|---|---|
cohere/north-mini-code | 256K | agentic, function_calling | $0 |
dots-studio/dots-3-note-preview | 512K | agentic, function_calling, json | $0 |
google/gemma-4-26b-a4b-it | 262K | agentic, function_calling, json | $0 |
google/gemma-4-31b-it | 262K | agentic, function_calling, json | $0 |
inclusionai/ling-3.0-flash-fin | 262K | agentic, function_calling | $0 |
inclusionai/ling-3.0-flash-sante | 262K | agentic, function_calling | $0 |
inclusionai/ling-3.0-flash | 262K | agentic, function_calling | $0 |
inclusionai/ling-3.0-tiny | 262K | agentic, function_calling | $0 |
liquid/lfm-2.5-2.6b | 66K | agentic, function_calling, json | $0 |
minimax/minimax-m2.7 | 197K | agentic, function_calling, json | $0 |
minimax/minimax-m3 | 1M | agentic, function_calling, json | $0 |
nvidia/nemotron-3-nano-30b-a3b | 256K | agentic, function_calling | $0 |
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning | 256K | agentic, function_calling | $0 |
nvidia/nemotron-3-super-120b-a12b | 262K | agentic, function_calling, json | $0 |
nvidia/nemotron-3-ultra-550b-a55b | 1M | agentic, function_calling | $0 |
nvidia/nemotron-3.5-lightning | 1M | agentic, function_calling | $0 |
nvidia/nemotron-nano-12b-v2-vl | 128K | agentic, function_calling | $0 |
nvidia/nemotron-nano-9b-v2 | 128K | agentic, function_calling, json | $0 |
openai/gpt-oss-20b | 131K | agentic, function_calling, json | $0 |
poolside/laguna-m.1 | 262K | agentic, function_calling | $0 |
poolside/laguna-s-2.1 | 262K | agentic, function_calling | $0 |
poolside/laguna-xs-2.1 | 262K | agentic, function_calling | $0 |
thinkingmachines/inkling-small | 1M | agentic, function_calling | $0 |
thinkingmachines/inkling | 1M | agentic, function_calling | $0 |
z-ai/glm-5.2 | 256K | agentic, function_calling, json | $0 |
Why this is safe to use in production
"Free model" usually means "you get what you pay for": silent tool-calling failures, malformed JSON, empty responses. Flow AI treats free capacity as supply to be verified, not trusted. Every candidate must pass a canary that exercises actual function-calling and output contracts before it can serve traffic, and the router's completion-aware cascade judges every response — a free model that fails your task escalates to a stronger model automatically, and rejected attempts aren't charged.
Try it in the playground, or read the docs — free models are already part of auto.