HomeNews

Edge0-145B: Open-Source 145B On-Device Model at 29GB

Kamran
Kamran
•
Nov 13, 2026 8:10 AM
•
0
 min read
Select Emergent as your Preferred news source
Edge0-145B: Open-Source 145B On-Device Model at 29GB

💡 TL;DR

  • Edge0 released Edge0-145B, the first open-source 145B-parameter model running fully on personal laptops with 29GB peak memory.
  • The model achieves 80+ tokens per second generation speed and scores 86.52 on Terminal-Bench 2.1, exceeding DeepSeek-V4-Flash.
  • Built with latent-MoE architecture, Edge0-145B activates only 10B parameters per step while supporting 128K to 256K context windows.

Edge0 has released Edge0-145B and Edge0-120B, marking a significant milestone in local AI computing. Officially launched on November 12, 2026, Edge0-145B is the first open-source 145-billion-parameter model designed to run entirely on personal computers without cloud dependencies. The model achieves 80+ tokens per second generation speed while consuming just 29GB of peak memory on a 48GB laptop, bringing frontier-level capabilities to consumer hardware.

Architecture and Performance Metrics

Edge0-145B employs a latent mixture-of-experts (MoE) design that activates approximately 10 billion parameters per inference step, despite having 145 billion total parameters. This architecture shares large matrices across 64 sparse experts, with each expert group sharing two large matrices while retaining small parameter sets for specialization. The model processes prompts at 1,100+ tokens per second and supports context windows ranging from 128K to 256K tokens.

The inference framework pipelines GPU memory, system memory, and SSD access across three storage tiers. Frequently accessed experts remain in GPU memory, while an expert library of approximately 40GB sits in system memory. Vocabulary data of roughly 48GB resides on SSD and loads on demand. This design combines CPU and GPU computation, using speculative decoding to verify multiple candidate tokens simultaneously and amortize weight-reading costs.

Benchmark Results and Competitive Position

Edge0-145B demonstrates competitive performance across multiple evaluation domains:

  • Terminal-Bench 2.1: 86.52, exceeding DeepSeek-V4-Flash (82.7) and Qwen3.8-Flash-Next (83.3)
  • GPQA-Diamond: 90.4, reflecting strong knowledge capabilities
  • IFBench: 80.6, surpassing Claude-Opus-4.6 (62.5) in instruction following
  • AA-LCR: 79.3 versus Claude's 78.0
  • DeepSWE 1.1: Shows 53.1% gain over Qwen3.8-27B baseline

The model achieves a 72.8 mean score across six benchmarks while using approximately 40% of the estimated memory footprint required for comparable cloud models at standard 4-bit quantization. This efficiency enables deployment on consumer laptops with 48GB RAM, previously insufficient for models of this capability level.

Agentic Capabilities and Workflow Integration

Edge0-145B supports complex multi-step workflows that combine planning, code generation, tool execution, debugging, and visual verification. The model can process natural-language requests that initiate entire local pipelines without external API calls. In testing, Edge0-145B produced educational videos about physics concepts, handling research, scriptwriting, asset generation, and editing within an hour. Manual production of equivalent work typically requires a team of specialists and a month of effort.

The model constructed a 3D Manhattan scene in Blender from text and reference images, iterating on materials, lighting, and rendered output through visual inspection and code revision. It also generated playable games and built an original medieval capital for a fantasy RPG with game-trailer-quality rendering. These workflows eliminate cloud-model token fees and keep sensitive code and data on local hardware.

Release and Availability

Edge0 open-sourced the full inference stack, including model weights, framework code, and deployment tools. Developers can reproduce results, fine-tune models, or integrate the technology into products. The release follows Edge0's earlier Edge0-35B phone model, which reported 37+ tokens per second on an iPhone 18 Pro Max using approximately 2.5GB memory. Edge0-145B scales to roughly four times the parameter count while maintaining practical resource consumption on laptop hardware.

What This Means

Edge0-145B challenges the assumption that frontier-level AI requires cloud infrastructure. By delivering 145-billion-parameter performance within consumer laptop constraints, the model enables privacy-preserving, cost-effective, and offline-capable AI workflows. The open-source release provides developers with tools to build local-first applications that previously required cloud dependencies. This shift may accelerate adoption of on-device AI for sensitive domains where data residency, latency, or connectivity constraints limit cloud solutions.

About the writer

Kamran Alam is an Engineer at Emergent, where he builds the core platform powering AI-driven software creation. He previously co-founded Jeevam Health (YC S20) as CTO and holds a B.Tech. from IIT Roorkee.

Start Building
on Emergent today
Try Emergent