Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-10 · high
The open-source framework offloads experts to SSD and predicts routing ahead of time, trading 4 points of accuracy for an 8x memory cut on Apple Silicon.
Why now: Edge0-AI published the repository on September 8, and its star count went from 269 to 583 within two days.
Edge0-AI/Edge0 packages SSD expert offload, Recover-LoRA and prerouter routing prediction into an MLX-backed framework built for Apple Silicon.
Its flagship model, edge0-35b, is a 4-bit, 40-layer, 256-expert build on Qwen3.5-MoE 35B-A3B. It holds peak active memory around 2.9 GB during inference, against a 23 GB on-disk checkpoint. That puts a 35-billion-parameter model within reach of a laptop instead of a rack of GPUs.
Most of the model's experts stay on SSD. Edge0 predicts ahead of time which experts a token will need and prefetches them, so disk reads don't stall generation.
The cost shows up in the benchmarks. edge0-35b averages 79.2 across AIME 2026, HumanEval, GPQA-Diamond, MMLU-Pro and IFBench, against 83.2 for the fp16 base model. Four points down for eight times less memory.
The repo reserves a CUDA backend directory but hasn't filled it in. Right now this only runs on Apple Silicon, and the repo doesn't say when, or whether, Nvidia support arrives.
Each link below shares sources, entities, or timing with this story.
Claims up to 2x faster generation, and adds fine-tuning of both MoE models on text or image datasets on Apple Silicon via MLX (release). Follow-up turns in long Qwen chats on Mac are reported up to 30x faster, MLX models now use full context size, and GLM-5.3 MLX fine-tunes ex...
Ollama cut v0.34.0-rc1 on September 5 at 23:49 UTC, and the headline item changes the shape of the local-versus-hosted decision rather than the performance of either side: Ollama-hosted open models can be selected directly inside ChatGPT Desktop, with setup driven from the Oll...
AlexsJones/llmfit released v1.1.10 today, adding RamaLama runtime discovery to its MCP server, the Qwen3.8 model family and MiniMax M3 vision capability exposure (GitHub). It also merged 32 MLX benchmark results on an Apple M4 Pro, the project's first MLX entries, giving an ap...
For about a year, "run your agent locally" meant accepting a model that couldn't reliably call a tool twice in a row. That excuse is gone. Meta Superintelligence Labs published Muse Glimmer today: a 29.6B dense causal transformer, 52 layers, 6,656 hidden dim, with a ~1.8B ViT-...
3B active parameters, beats Qwen3.5-35B-A3B on AIME 2025 (92.4 vs 91.9), LiveCodeBench v6 (87.2 vs 74.6), and surpasses the larger Nemotron-3-Super-120B. Available on Ollama and HuggingFace under open license. Source
Google released Gemma 4 on April 2 with four model variants: E2B, E4B, 26B MoE, and 31B Dense. The license change is the first thing worth noting. Every previous Gemma had restrictions that made lawyers nervous. Gemma 4 is Apache 2.0. Full stop. Use it in any product, any way...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.