Edge0 runs a 35B model in 3 GB by paging experts off your SSD
The Apache-2.0 MoE keeps weights on disk and pulls in only what it needs — plus DeepSeek's MIT flash model at $0.15 and Tencent's open speech model.
Copy markdown
A 35B model that fits in 3 GB of RAM
Edge0-35B-A3B streams its experts off your SSD instead of loading all the weights, holding under 3 GB of active memory while a prerouter prefetches the next experts one step ahead. It's Apache 2.0 and runs ~15-18 tokens/sec on Apple Silicon — the catch is it's MLX-only for now, other backends still on the roadmap.
DeepSeek's MIT flash model took over your old endpoint
V4.1 Flash is open MIT weights at $0.15 per million input tokens (off-peak), a 1M-token context, and it matches Claude Opus-5.0 on DeepSWE (74.2). Heads-up: calls to the old deepseek-v4-flash now redirect to it and bill at its price — switch your code to deepseek-flash.
Tencent's AuK does TTS and speech recognition in one 1.5B model
The open-weight AuK handles zero-shot text-to-speech with voice cloning and speech-to-text from a single 1.5B checkpoint on Hugging Face. That's one small dependency to add voice in and out of an app, instead of wiring up two separate models.
Elsewhere: MiniMax ships H3 video weights, but not to the West
MiniMax's ~42.5 GB open video model does text-to-video and image-to-video, but its license shuts out users in the US, EU, and two other regions — check it before you build on the weights.