Edge0 open-sources a 145B model that fits on a laptop in 29GB

A developer paying a hosted model for every request now has an open-source option. The whole job runs on the laptop. Edge0 released Edge0-145B on 12 November 2026. The licence is open source. The company says the model uses 29GB peak memory. It decodes above 80 tokens per second on a single laptop.

The release page carries Edge0's own description. The company calls Edge0-145B the world's first open-source on-device model at that scale. Edge0 also open-sourced a smaller Edge0-120B. A unified inference framework and deployment tools came with the models.

What the hardware looks like

Edge0 lists day-one support for macOS, Windows and Linux. The hardware list covers Apple M-series chips and NVIDIA platforms, including RTX Spark, RTX 4090 and RTX 5090. The company states 120 watts of typical power draw. It states 1,100 tokens per second on prompt prefill and 80 tokens per second on decode. Context length runs from 128,000 to 256,000 tokens.

The architecture details are Edge0's own. The company says it shares matrices across experts. That cuts the weights the laptop must store and move. The expert count stays the same. Training happens in the on-device architecture from the start, so deployment limits shape the design. Reading and compute overlap across three storage tiers. The GPU runs cached experts while the CPU handles the rest.

Edge0's benchmark claims

Every number here is Edge0's own, read from its release page. On DeepSWE 1.1 for coding, Edge0-145B scores 64.6. The company puts Qwen3.8-Flash-Next at 58.7 and DeepSeek-V4-Flash-0731 at 54.4 in the same table. On Terminal-Bench 2.1, Edge0 reports 86.52, with Qwen3.8-Flash-Next at 83.3, DeepSeek at 82.7 and Claude at 78.2. On GPQA-Diamond, Edge0 reports 90.4, with DeepSeek at 90.8 and Claude-Opus-4.6 Max at 91.3. On IFBench, Edge0 reports 80.6, with Claude-Opus-4.6 Max at 62.5. On AA-LCR, Edge0 reports 79.3, with Claude at 78.0.

None of these numbers are independent. The release page is the vendor's own card, and the owner has not yet published a third-party benchmark. A team should treat the figures as what to test against on its own task, not as settled rankings.

Where to test Edge0-145B first

A team runs retrieval, drafting, review or classification on a cloud model today. It can now test the same workload on hardware the engineers already own. The release removes a vendor subscription in one step. It also removes the handling question for private data. The cost is a few hours of engineer time. Run the model on the real task. Compare its output to what the cloud model returns. The practical reading on these claims is not the leaderboard order. It is that a 145-billion-parameter open model now fits in the memory of a workstation laptop.

Source

Introducing Edge0-145B, Edge0.