Edge0 fits a 35B model in under 3GB, streaming experts off SSD
Apache-2.0, ~15 tok/s on an M4 Pro, 3.9-pt quality drop. Also new: Nex-N2.5 open computer-use agents and MiniCPM5-2B, a 2B beating Qwen3.5-4B.
Copy markdown
A 35B-class model on a phone, in ~3GB
Edge0-35B-A3B keeps peak active memory under 3 GiB (2.9 measured) while running a 35B-parameter Qwen3.5-MoE, no GPU required. It's Apache-2.0 with weights on Hugging Face and GitHub — the first genuinely local option at this size class.
How: experts streamed off SSD, 4-bit + LoRA
Only ~3B of 35B params fire per token; the rest stay on disk and stream in on demand, so RAM holds just the active weights. A trained prerouter predicts the next token's experts for up to 59% higher decode, and 4-bit weights with Recover-LoRA adapters keep quality within 3.9 points of full precision.
Speed, and the catch
Expect ~15 tok/s decode (113–140 prefill) on an M4 Pro, and you'll want fast NVMe/flash for the streaming to pay off. Today it's MLX / Apple-Silicon only — Windows, Linux and NVIDIA are roadmap — and this preview is weak at tool use and long-horizon agent work.
Build this weekend: a fully local reasoning app
A private, offline chat or reasoning tool on a laptop or iPhone with no API bill and no data leaving the device — and you can batch-serve several LoRA variants off one base. Best for chat, math and code reasoning today, not autonomous agents.
Elsewhere: Nex-N2.5 open computer-use agents
Nex-AGI shipped open-weight Nex-N2.5 Mini (35B) and Pro (397B) — multimodal models trained to drive a screen (Pro hits 82.2 on OSWorld-Verified) — plus a 1.6T text-only Max. Mini and Pro are live on OpenRouter, with weights on Hugging Face and an sglang Docker for self-hosting.
Elsewhere: MiniCPM5-2B beats a 4B
OpenBMB's 2.52B MiniCPM5-2B averages 53.9 across 34 benchmarks — past Qwen3.5-4B's 51.1 — with 69.1 on LiveCodeBench and a 128K context. Apache-2.0, standard Llama arch, runs in llama.cpp, Ollama, LM Studio and MLX.