Open source alternatives to Gemini 3.5 Flash

By GitHub stars it is Step-Audio-R1, with 702 stars and a commit on 29 Apr. Choose Step-Audio-R1 when audio understanding must run on your own servers instead of sending recordings to Google.

Yes. The code of all 3 is public under Apache 2.0 or MIT, so you can use them with no licence fee. Running one yourself costs the server and your time to look after it, and some makers also sell a hosted plan.

Start from the job you use Gemini 3.5 Flash for. Choose DeepSeek V4.1 Flash when cost matters most; it charges $0.30 / $1.20 per million tokens at peak hours and half that off-peak, with a 1M context and up to 384K output, though it lists image input but not audio or video. Choose Bonsai 27B when you want a 27B model compressed to 1-bit or ternary weights that runs locally on modest hardware.

All 3: Step-Audio-R1, Bonsai 27B and DeepSeek V4.1 Flash. Each tool's page says what it needs to run.

0 of the 1 we track had a commit in the last 30 days. The quietest is Step-Audio-R1, last changed on 29 Apr. A recent commit shows the project is kept up, not that it will stay that way.

Apache 2.0 (2), MIT (1). MIT, BSD and Apache let you change and reuse the code freely.

The hosting. Gemini 3.5 Flash's maker runs the servers for you; with an open source tool the updates, backups, security fixes and support are yours, unless you take the maker's hosted plan. Features differ too, so check what each one imports and exports before you move your data.