Edge0: a 35B MoE that runs in under 3GB of RAM, no GPU
An Apache-2.0 preview streams experts off SSD, trading ~4 points of quality to run offline — plus OpenAI's full-duplex voice hits the API at $0.05/min.
Copy markdown
Edge0-35B-A3B: a 35B model in a phone's worth of RAM
The preview keeps all 35B parameters (4 of 256 experts active per token) on SSD and streams only what each token routes, peaking at 2.9GB RAM on a Mac mini M4 Pro. Weights are 4-bit (19.6GB on disk), Apache 2.0, and text-only — you own it outright.
The trick: SSD expert offload plus a look-ahead prerouter
A trained prerouter predicts which experts the next token will need and prefetches them, overlapping storage reads with compute for up to 59% higher decode throughput. Real speed lands around 15-18 tokens/sec decode and 113-140 tokens/sec prefill — usable, not blazing.
The catch: preview quality, and weak at agents
Against an FP16 Qwen3.5-MoE baseline it averages ~3.9 points lower (79.2 vs 83.2; AIME 86.6 vs 92.7, HumanEval 90.9 vs 95.1). The model card is blunt that tool use, multi-step planning, and long-horizon autonomy are currently weak — treat it as a chat and reasoning model, not an agent.
Build this weekend: a private assistant with no GPU bill
A capable 35B reasoner now fits on a laptop or Mac mini with RAM to spare, fully offline. Because it serves multiple LoRA adapters off one base model, you can host several task-specialized assistants from a single ~3GB footprint.
Elsewhere: OpenAI's full-duplex voice, $0.05/min in the API
GPT-Live-1 reached the API with native full-duplex speech and telephony, billed at $0.05 per minute. It collapses real-time voice agents into one call — no more wiring separate speech-to-text, LLM, and text-to-speech stages together.